Search

We present a deep learning-based framework to design and quantify point-of-care sensors. As a use-case, we demonstrated a low-cost and rapid paper-based vertical flow assay (VFA) for high sensitivity C-Reactive Protein (hsCRP) testing, commonly used for assessing risk of cardio-vascular disease (CVD). A machine learning-based framework was developed to (1) determine an optimal configuration of immunoreaction spots and conditions, spatially-multiplexed on a sensing membrane, and (2) to accurately infer target analyte concentration. Using a custom-designed handheld VFA reader, a clinical study with 85 human samples showed a competitive coefficient-of-variation of 11.2% and linearity of R2 = 0.95 among blindly-tested VFAs in the hsCRP range (i.e., 0–10 mg/L). We also demonstrated a mitigation of the hook-effect due to the multiplexed immunoreactions on the sensing membrane. This paper-based computational VFA could expand access to CVD testing, and the presented framework can be broadly used to design cost-effective and mobile point-of-care sensors. [ABSTRACT FROM AUTHOR]

This paper focuses on how regulatory bodies respond to artificial intelligence (AI)-enabled medical devices. To achieve this, we present a comparative overview of the United States (USA), European Union (EU), and China. Our search in the governmental database identified 59 AI medical devices approved in China as of July 2023. In comparison to the rules-based regulatory approach in China, the approaches in the USA and EU are more standards-oriented. [ABSTRACT FROM AUTHOR]

Advancements in deep learning and computer vision provide promising solutions for medical image analysis, potentially improving healthcare and patient outcomes. However, the prevailing paradigm of training deep learning models requires large quantities of labeled training data, which is both time-consuming and cost-prohibitive to curate for medical images. Self-supervised learning has the potential to make significant contributions to the development of robust medical imaging models through its ability to learn useful insights from copious medical datasets without labels. In this review, we provide consistent descriptions of different self-supervised learning strategies and compose a systematic review of papers published between 2012 and 2022 on PubMed, Scopus, and ArXiv that applied self-supervised learning to medical imaging classification. We screened a total of 412 relevant studies and included 79 papers for data extraction and analysis. With this comprehensive effort, we synthesize the collective knowledge of prior work and provide implementation guidelines for future researchers interested in applying self-supervised learning to their development of medical imaging classification models. [ABSTRACT FROM AUTHOR]

Machine learning and deep learning are two subsets of artificial intelligence that involve teaching computers to learn and make decisions from any sort of data. Most recent developments in artificial intelligence are coming from deep learning, which has proven revolutionary in almost all fields, from computer vision to health sciences. The effects of deep learning in medicine have changed the conventional ways of clinical application significantly. Although some sub-fields of medicine, such as pediatrics, have been relatively slow in receiving the critical benefits of deep learning, related research in pediatrics has started to accumulate to a significant level, too. Hence, in this paper, we review recently developed machine learning and deep learning-based solutions for neonatology applications. We systematically evaluate the roles of both classical machine learning and deep learning in neonatology applications, define the methodologies, including algorithmic developments, and describe the remaining challenges in the assessment of neonatal diseases by using PRISMA 2020 guidelines. To date, the primary areas of focus in neonatology regarding AI applications have included survival analysis, neuroimaging, analysis of vital parameters and biosignals, and retinopathy of prematurity diagnosis. We have categorically summarized 106 research articles from 1996 to 2022 and discussed their pros and cons, respectively. In this systematic review, we aimed to further enhance the comprehensiveness of the study. We also discuss possible directions for new AI models and the future of neonatology with the rising power of AI, suggesting roadmaps for the integration of AI into neonatal intensive care units. [ABSTRACT FROM AUTHOR]

Nocturnal scratch is one major factor leading to impaired quality of life in atopic dermatitis (AD) patients. Therefore, objectively quantifying nocturnal scratch events aids in assessing the disease state, treatment effect, and AD patients' quality of life. In this paper, we describe the use of actigraphy, highly predictive topological features, and a model-ensembling approach to develop an assessment of nocturnal scratch events by measuring scratch duration and intensity. Our assessment is tested in a clinical setting against the ground truth obtained from video recordings. The new approach addresses unmet challenges in existing studies, such as the lack of generalizability to real-world applications, the failure to capture finger scratches, and the limitations in the evaluation due to imbalanced data in the current literature. Furthermore, the performance evaluation shows agreement between derived digital endpoints and the video annotation ground truth, as well as patient-reported outcomes, which demonstrated the validity of the new assessment of nocturnal scratch. [ABSTRACT FROM AUTHOR]

Data privacy mechanisms are essential for rapidly scaling medical training databases to capture the heterogeneity of patient data distributions toward robust and generalizable machine learning systems. In the current COVID-19 pandemic, a major focus of artificial intelligence (AI) is interpreting chest CT, which can be readily used in the assessment and management of the disease. This paper demonstrates the feasibility of a federated learning method for detecting COVID-19 related CT abnormalities with external validation on patients from a multinational study. We recruited 132 patients from seven multinational different centers, with three internal hospitals from Hong Kong for training and testing, and four external, independent datasets from Mainland China and Germany, for validating model generalizability. We also conducted case studies on longitudinal scans for automated estimation of lesion burden for hospitalized COVID-19 patients. We explore the federated learning algorithms to develop a privacy-preserving AI model for COVID-19 medical image diagnosis with good generalization capability on unseen multinational datasets. Federated learning could provide an effective mechanism during pandemics to rapidly develop clinically useful AI across institutions and countries overcoming the burden of central aggregation of large amounts of sensitive data. [ABSTRACT FROM AUTHOR]

An amendment to this paper has been published and can be accessed via a link at the top of the paper. [ABSTRACT FROM AUTHOR]

A Correction to this paper has been published: https://doi.org/10.1038/s41746-020-00365-5. [ABSTRACT FROM AUTHOR]

CT is crucial for diagnosing chest diseases, with image quality affected by spatial resolution. Thick-slice CT remains prevalent in practice due to cost considerations, yet its coarse spatial resolution may hinder accurate diagnoses. Our multicenter study develops a deep learning synthetic model with Convolutional-Transformer hybrid encoder-decoder architecture for generating thin-slice CT from thick-slice CT on a single center (1576 participants) and access the synthetic CT on three cross-regional centers (1228 participants). The qualitative image quality of synthetic and real thin-slice CT is comparable (p = 0.16). Four radiologists' accuracy in diagnosing community-acquired pneumonia using synthetic thin-slice CT surpasses thick-slice CT (p < 0.05), and matches real thin-slice CT (p > 0.99). For lung nodule detection, sensitivity with thin-slice CT outperforms thick-slice CT (p < 0.001) and comparable to real thin-slice CT (p > 0.05). These findings indicate the potential of our model to generate high-quality synthetic thin-slice CT as a practical alternative when real thin-slice CT is preferred but unavailable. [ABSTRACT FROM AUTHOR]

Corneal transplantation is the primary treatment for irreversible corneal diseases, but due to limited donor availability, bioengineered corneal equivalents are being developed as a solution, with biocompatibility, structural integrity, and physical function considered key factors. Since conventional evaluation methods may not fully capture the complex properties of the cornea, there is a need for advanced imaging and assessment techniques. In this study, we proposed a deep learning-based automatic segmentation method for transplanted bioengineered corneal equivalents using optical coherence tomography to achieve a highly accurate evaluation of graft integrity and biocompatibility. Our method provides quantitative individual thickness values, detailed maps, and volume measurements of the bioengineered corneal equivalents, and has been validated through 14 days of monitoring. Based on the results, it is expected to have high clinical utility as a quantitative assessment method for human keratoplasties, including automatic opacity area segmentation and implanted graft part extraction, beyond animal studies. [ABSTRACT FROM AUTHOR]

Clinical trial matching is the task of identifying trials for which patients may be eligible. Typically, this task is labor-intensive and requires detailed verification of patient electronic health records (EHRs) against the stringent inclusion and exclusion criteria of clinical trials. This process also results in many patients missing out on potential therapeutic options. Recent advancements in Large Language Models (LLMs) have made automating patient-trial matching possible, as shown in multiple concurrent research studies. However, the current approaches are confined to constrained, often synthetic, datasets that do not adequately mirror the complexities encountered in real-world medical data. In this study, we present an end-to-end large-scale empirical evaluation of a clinical trial matching system and validate it using real-world EHRs. We perform comprehensive experiments with proprietary LLMs and our custom fine-tuned model called OncoLLM and show that OncoLLM outperforms GPT-3.5 and matches the performance of qualified medical doctors for clinical trial matching. [ABSTRACT FROM AUTHOR]

Increasing interest in applying large language models (LLMs) to medicine is due in part to their impressive performance on medical exam questions. However, these exams do not capture the complexity of real patient–doctor interactions because of factors like patient compliance, experience, and cognitive bias. We hypothesized that LLMs would produce less accurate responses when faced with clinically biased questions as compared to unbiased ones. To test this, we developed the BiasMedQA dataset, which consists of 1273 USMLE questions modified to replicate common clinically relevant cognitive biases. We assessed six LLMs on BiasMedQA and found that GPT-4 stood out for its resilience to bias, in contrast to Llama 2 70B-chat and PMC Llama 13B, which showed large drops in performance. Additionally, we introduced three bias mitigation strategies, which improved but did not fully restore accuracy. Our findings highlight the need to improve LLMs' robustness to cognitive biases, in order to achieve more reliable applications of LLMs in healthcare. [ABSTRACT FROM AUTHOR]

The success of deep learning (DL) relies heavily on training data from which DL models encapsulate information. Consequently, the development and deployment of DL models expose data to potential privacy breaches, which are particularly critical in data-sensitive contexts like medicine. We propose a new technique named DiffGuard that generates realistic and diverse synthetic medical images with annotations, even indistinguishable for experts, to replace real data for DL model training, which cuts off their direct connection and enhances privacy safety. We demonstrate that DiffGuard enhances privacy safety with much less data leakage and better resistance against privacy attacks on data and model. It also improves the accuracy and generalizability of DL models for segmentation and classification of mediastinal neoplasms in multi-center evaluation. We expect that our solution would enlighten the road to privacy-preserving DL for precision medicine, promote data and model sharing, and inspire more innovation on artificial-intelligence-generated-content technologies for medicine. [ABSTRACT FROM AUTHOR]

An amendment to this paper has been published and can be accessed via a link at the top of the paper. [ABSTRACT FROM AUTHOR]

Personal and family history of suicidal thoughts and behaviors (PSH and FSH, respectively) are significant risk factors associated with suicides. Research is limited in automatic identification of such data from clinical notes in Electronic Health Records. This study developed deep learning (DL) tools utilizing transformer models (Bio_ClinicalBERT and GatorTron) to detect PSH and FSH in clinical notes derived from three academic medical centers, and compared their performance with a rule-based natural language processing tool. For detecting PSH, the rule-based approach obtained an F1-score of 0.75 ± 0.07, while the Bio_ClinicalBERT and GatorTron DL tools scored 0.83 ± 0.09 and 0.84 ± 0.07, respectively. For detecting FSH, the rule-based approach achieved an F1-score of 0.69 ± 0.11, compared to 0.89 ± 0.10 for Bio_ClinicalBERT and 0.92 ± 0.07 for GatorTron. Across sites, the DL tools identified more than 80% of patients at elevated risk for suicide who remain undiagnosed and untreated. [ABSTRACT FROM AUTHOR]

Large language models (LLMs) hold great promise in summarizing medical evidence. Most recent studies focus on the application of proprietary LLMs. Using proprietary LLMs introduces multiple risk factors, including a lack of transparency and vendor dependency. While open-source LLMs allow better transparency and customization, their performance falls short compared to the proprietary ones. In this study, we investigated to what extent fine-tuning open-source LLMs can further improve their performance. Utilizing a benchmark dataset, MedReview, consisting of 8161 pairs of systematic reviews and summaries, we fine-tuned three broadly-used, open-sourced LLMs, namely PRIMERA, LongT5, and Llama-2. Overall, the performance of open-source models was all improved after fine-tuning. The performance of fine-tuned LongT5 is close to GPT-3.5 with zero-shot settings. Furthermore, smaller fine-tuned models sometimes even demonstrated superior performance compared to larger zero-shot models. The above trends of improvement were manifested in both a human evaluation and a larger-scale GPT4-simulated evaluation. [ABSTRACT FROM AUTHOR]

Distributed collaborative learning is a promising approach for building predictive models for privacy-sensitive biomedical images. Here, several data owners (clients) train a joint model without sharing their original data. However, concealed systematic biases can compromise model performance and fairness. This study presents MyThisYourThat (MyTH) approach, which adapts an interpretable prototypical part learning network to a distributed setting, enabling each client to visualize feature differences learned by others on their own image: comparing one client's 'This' with others' 'That'. Our setting demonstrates four clients collaboratively training two diagnostic classifiers on a benchmark X-ray dataset. Without data bias, the global model reaches 74.14% balanced accuracy for cardiomegaly and 74.08% for pleural effusion. We show that with systematic visual bias in one client, the performance of global models drops to near-random. We demonstrate how differences between local and global prototypes reveal biases and allow their visualization on each client's data without compromising privacy. [ABSTRACT FROM AUTHOR]

Clostridioides difficile infection (CDI) is a major cause of antibiotic-associated diarrhea and colitis. It is recognized as one of the most significant hospital-acquired infections. Although CDI can develop severe complications and spores of Clostridioides difficile can be transmitted by the fecal-oral route, CDI is occasionally overlooked in clinical settings. Thus, it is necessary to monitor high CDI risk groups, particularly those undergoing antibiotic treatment, to prevent complications and spread. We developed and validated a deep learning-based model to predict the occurrence of CDI within 28 days after starting antibiotic treatment using longitudinal electronic health records. For each patient, timelines of vital signs and laboratory tests with a 35-day monitoring period and a patient information vector consisting of age, sex, comorbidities, and medications were constructed. Our model achieved the prediction performance with an area under the receiver operating characteristic curve of 0.952 (95% CI: 0.932–0.973) in internal validation and 0.972 (95% CI: 0.968–0.975) in external validation. Platelet count and body temperature emerged as the most important features. The risk score, the output value of the model, exhibited a consistent increase in the CDI group, while the risk score in the non-CDI group either maintained its initial value or decreased. Using our CDI prediction model, high-risk patients requiring symptom monitoring can be identified. This could help reduce the underdiagnosis of CDI, thereby decreasing transmission and preventing complications. [ABSTRACT FROM AUTHOR]

Calcification has significant influence over cardiovascular diseases and interventions. Detailed characterization of calcification is thus desired for predictive modeling, but calcium deposits on cardiovascular structures are still often manually reconstructed for physics-driven simulations. This poses a major bottleneck for large-scale adoption of computational simulations for research or clinical use. To address this, we propose an end-to-end automated image-to-mesh algorithm that enables robust incorporation of patient-specific calcification onto a given cardiovascular tissue mesh. The algorithm provides a substantial speed-up from several hours of manual meshing to ~1 min of automated computation, and it solves an important problem that cannot be addressed with recent template-based meshing techniques. We validated our final calcified tissue meshes with extensive simulations, demonstrating our ability to accurately model patient-specific aortic stenosis and Transcatheter Aortic Valve Replacement. Our method may serve as an important tool for accelerating the development and usage of personalized cardiovascular biomechanics. [ABSTRACT FROM AUTHOR]

Parkinson's disease (PD) is a serious neurodegenerative disorder marked by significant clinical and progression heterogeneity. This study aimed at addressing heterogeneity of PD through integrative analysis of various data modalities. We analyzed clinical progression data (≥5 years) of individuals with de novo PD using machine learning and deep learning, to characterize individuals' phenotypic progression trajectories for PD subtyping. We discovered three pace subtypes of PD exhibiting distinct progression patterns: the Inching Pace subtype (PD-I) with mild baseline severity and mild progression speed; the Moderate Pace subtype (PD-M) with mild baseline severity but advancing at a moderate progression rate; and the Rapid Pace subtype (PD-R) with the most rapid symptom progression rate. We found cerebrospinal fluid P-tau/α-synuclein ratio and atrophy in certain brain regions as potential markers of these subtypes. Analyses of genetic and transcriptomic profiles with network-based approaches identified molecular modules associated with each subtype. For instance, the PD-R-specific module suggested STAT3, FYN, BECN1, APOA1, NEDD4, and GATA2 as potential driver genes of PD-R. It also suggested neuroinflammation, oxidative stress, metabolism, PI3K/AKT, and angiogenesis pathways as potential drivers for rapid PD progression (i.e., PD-R). Moreover, we identified repurposable drug candidates by targeting these subtype-specific molecular modules using network-based approach and cell line drug-gene signature data. We further estimated their treatment effects using two large-scale real-world patient databases; the real-world evidence we gained highlighted the potential of metformin in ameliorating PD progression. In conclusion, this work helps better understand clinical and pathophysiological complexity of PD progression and accelerate precision medicine. [ABSTRACT FROM AUTHOR]

Sleep is essential to life. Accurate measurement and classification of sleep/wake and sleep stages is important in clinical studies for sleep disorder diagnoses and in the interpretation of data from consumer devices for monitoring physical and mental well-being. Existing non-polysomnography sleep classification techniques mainly rely on heuristic methods developed in relatively small cohorts. Thus, we aimed to establish the accuracy of wrist-worn accelerometers for sleep stage classification and subsequently describe the association between sleep duration and efficiency (proportion of total time asleep when in bed) with mortality outcomes. We developed a self-supervised deep neural network for sleep stage classification using concurrent laboratory-based polysomnography and accelerometry. After exclusion, 1113 participant nights of data were used for training. The difference between polysomnography and the model classifications on the external validation was 48.2 min (95% limits of agreement (LoA): −50.3 to 146.8 min) for total sleep duration, −17.1 min for REM duration (95% LoA: −56.7 to 91.0 min) and 31.1 min (95% LoA: −67.3 to 129.5 min) for NREM duration. The sleep classifier was deployed in the UK Biobank with ~100,000 participants to study the association of sleep duration and sleep efficiency with all-cause mortality. Among 66,262 UK Biobank participants, 1644 mortality events were observed. Short sleepers (<6 h) had a higher risk of mortality compared to participants with normal sleep duration 6–7.9 h, regardless of whether they had low sleep efficiency (Hazard ratios (HRs): 1.36; 95% confidence intervals (CIs): 1.18 to 1.58) or high sleep efficiency (HRs: 1.29; 95% CIs: 1.04–1.61). Deep-learning-based sleep classification using accelerometers has a fair to moderate agreement with polysomnography. Our findings suggest that having short overnight sleep confers mortality risk irrespective of sleep continuity. [ABSTRACT FROM AUTHOR]

AI-enabled ECGs have previously been shown to accurately predict patient sex in adults and correlate with sex hormone levels. We aimed to test the ability of AI-enabled ECGs to predict sex in the pediatric population and study the influence of pubertal development. AI-enabled ECG models were created using a convolutional neural network trained on pediatric 10-second, 12-lead ECGs. The first model was trained de novo using pediatric data. The second model used transfer learning from a previously validated adult data-derived algorithm. We analyzed the first ECG from 90,133 unique pediatric patients (aged ≤18 years) recorded between 1987–2022, and divided the cohort into training, validation, and testing datasets. Subgroup analysis was performed on prepubertal (0–7 years), peripubertal (8–14 years), and postpubertal (15–18 years) patients. The cohort was 46.7% male, with 21,678 prepubertal, 26,740 peripubertal, and 41,715 postpubertal children. The de novo pediatric model demonstrated 81% accuracy and an area under the curve (AUC) of 0.91. Model sensitivity was 0.79, specificity was 0.83, positive predicted value was 0.84, and the negative predicted value was 0.78, for the entire test cohort. The model's discriminatory ability was highest in postpubertal (AUC = 0.98), lower in the peripubertal age group (AUC = 0.91), and poor in the prepubertal age group (AUC = 0.67). There was no significant performance difference observed between the transfer learning and de novo models. AI-enabled interpretation of ECG can estimate sex in peripubertal and postpubertal children with high accuracy. [ABSTRACT FROM AUTHOR]

Early-life exposure to stress results in significantly increased risk of neurodevelopmental impairments with potential long-term effects into childhood and even adulthood. As a crucial step towards monitoring neonatal stress in neonatal intensive care units (NICUs), our study aims to quantify the duration, frequency, and physiological responses of care manipulation activities, based on bedside videos and physiological signals. Leveraging 289 h of video recordings and physiological data within 330 sessions collected from 27 neonates in 2 NICUs, we develop and evaluate a deep learning method to detect manipulation activities from the video, to estimate their duration and frequency, and to further integrate physiological signals for assessing their responses. With a 13.8% relative error tolerance for activity duration and frequency, our results were statistically equivalent to human annotations. Further, our method proved effective for estimating short-term physiological responses, for detecting activities with marked physiological deviations, and for quantifying the neonatal infant stressor scale scores. [ABSTRACT FROM AUTHOR]

Dystonia is a neurological movement disorder characterised by abnormal involuntary movements and postures, particularly affecting the head and neck. However, current clinical assessment methods for dystonia rely on simplified rating scales which lack the ability to capture the intricate spatiotemporal features of dystonic phenomena, hindering clinical management and limiting understanding of the underlying neurobiology. To address this, we developed a visual perceptive deep learning framework that utilizes standard clinical videos to comprehensively evaluate and quantify disease states and the impact of therapeutic interventions, specifically deep brain stimulation. This framework overcomes the limitations of traditional rating scales and offers an efficient and accurate method that is rater-independent for evaluating and monitoring dystonia patients. To evaluate the framework, we leveraged semi-standardized clinical video data collected in three retrospective, longitudinal cohort studies across seven academic centres. We extracted static head angle excursions for clinical validation and derived kinematic variables reflecting naturalistic head dynamics to predict dystonia severity, subtype, and neuromodulation effects. The framework was also applied to a fully independent cohort of generalised dystonia patients for comparison between dystonia sub-types. Computer vision-derived measurements of head angle excursions showed a strong correlation with clinically assigned scores. Across comparisons, we identified consistent kinematic features from full video assessments encoding information critical to disease severity, subtype, and effects of neural circuit interventions, independent of static head angle deviations used in scoring. Our visual perceptive machine learning framework reveals kinematic pathosignatures of dystonia, potentially augmenting clinical management, facilitating scientific translation, and informing personalized precision neurology approaches. [ABSTRACT FROM AUTHOR]

Apnea and hypopnea are common sleep disorders characterized by the obstruction of the airways. Polysomnography (PSG) is a sleep study typically used to compute the Apnea-Hypopnea Index (AHI), the number of times a person has apnea or certain types of hypopnea per hour of sleep, and diagnose the severity of the sleep disorder. Early detection and treatment of apnea can significantly reduce morbidity and mortality. However, long-term PSG monitoring is unfeasible as it is costly and uncomfortable for patients. To address these issues, we propose a method, named DRIVEN, to estimate AHI at home from wearable devices and detect when apnea, hypopnea, and periods of wakefulness occur throughout the night. The method can therefore assist physicians in diagnosing the severity of apneas. Patients can wear a single sensor or a combination of sensors that can be easily measured at home: abdominal movement, thoracic movement, or pulse oximetry. For example, using only two sensors, DRIVEN correctly classifies 72.4% of all test patients into one of the four AHI classes, with 99.3% either correctly classified or placed one class away from the true one. This is a reasonable trade-off between the model's performance and the patient's comfort. We use publicly available data from three large sleep studies with a total of 14,370 recordings. DRIVEN consists of a combination of deep convolutional neural networks and a light-gradient-boost machine for classification. It can be implemented for automatic estimation of AHI in unsupervised long-term home monitoring systems, reducing costs to healthcare systems and improving patient care. [ABSTRACT FROM AUTHOR]

Healthcare datasets are becoming larger and more complex, necessitating the development of accurate and generalizable AI models for medical applications. Unstructured datasets, including medical imaging, electrocardiograms, and natural language data, are gaining attention with advancements in deep convolutional neural networks and large language models. However, estimating the generalizability of these models to new healthcare settings without extensive validation on external data remains challenging. In experiments across 13 datasets including X-rays, CTs, ECGs, clinical discharge summaries, and lung auscultation data, our results demonstrate that model performance is frequently overestimated by up to 20% on average due to shortcut learning of hidden data acquisition biases (DAB). Shortcut learning refers to a phenomenon in which an AI model learns to solve a task based on spurious correlations present in the data as opposed to features directly related to the task itself. We propose an open source, bias-corrected external accuracy estimate, PEst, that better estimates external accuracy to within 4% on average by measuring and calibrating for DAB-induced shortcut learning. [ABSTRACT FROM AUTHOR]

Distribution shifts remain a problem for the safe application of regulated medical AI systems, and may impact their real-world performance if undetected. Postmarket shifts can occur for example if algorithms developed on data from various acquisition settings and a heterogeneous population are predominantly applied in hospitals with lower quality data acquisition or other centre-specific acquisition factors, or where some ethnicities are over-represented. Therefore, distribution shift detection could be important for monitoring AI-based medical products during postmarket surveillance. We implemented and evaluated three deep-learning based shift detection techniques (classifier-based, deep kernel, and multiple univariate kolmogorov-smirnov tests) on simulated shifts in a dataset of 130'486 retinal images. We trained a deep learning classifier for diabetic retinopathy grading. We then simulated population shifts by changing the prevalence of patients' sex, ethnicity, and co-morbidities, and example acquisition shifts by changes in image quality. We observed classification subgroup performance disparities w.r.t. image quality, patient sex, ethnicity and co-morbidity presence. The sensitivity at detecting referable diabetic retinopathy ranged from 0.50 to 0.79 for different ethnicities. This motivates the need for detecting shifts after deployment. Classifier-based tests performed best overall, with perfect detection rates for quality and co-morbidity subgroup shifts at a sample size of 1000. It was the only method to detect shifts in patient sex, but required large sample sizes ( > 3 0 ′ 000 ). All methods identified easier-to-detect out-of-distribution shifts with small (≤300) sample sizes. We conclude that effective tools exist for detecting clinically relevant distribution shifts. In particular classifier-based tests can be easily implemented components in the post-market surveillance strategy of medical device manufacturers. [ABSTRACT FROM AUTHOR]

Spectral-domain optical coherence tomography (SDOCT) is the gold standard of imaging the eye in clinics. Penetration depth with such devices is, however, limited and visualization of the choroid, which is essential for diagnosing chorioretinal disease, remains limited. Whereas swept-source OCT (SSOCT) devices allow for visualization of the choroid these instruments are expensive and availability in praxis is limited. We present an artificial intelligence (AI)-based solution to enhance the visualization of the choroid in OCT scans and allow for quantitative measurements of choroidal metrics using generative deep learning (DL). Synthetically enhanced SDOCT B-scans with improved choroidal visibility were generated, leveraging matching images to learn deep anatomical features during the training. Using a single-center tertiary eye care institution cohort comprising a total of 362 SDOCT-SSOCT paired subjects, we trained our model with 150,784 images from 410 healthy, 192 glaucoma, and 133 diabetic retinopathy eyes. An independent external test dataset of 37,376 images from 146 eyes was deployed to assess the authenticity and quality of the synthetically enhanced SDOCT images. Experts' ability to differentiate real versus synthetic images was poor (47.5% accuracy). Measurements of choroidal thickness, area, volume, and vascularity index, from the reference SSOCT and synthetically enhanced SDOCT, showed high Pearson's correlations of 0.97 [95% CI: 0.96–0.98], 0.97 [0.95–0.98], 0.95 [0.92–0.98], and 0.87 [0.83–0.91], with intra-class correlation values of 0.99 [0.98–0.99], 0.98 [0.98–0.99], and 0.95 [0.96–0.98], 0.93 [0.91–0.95], respectively. Thus, our DL generative model successfully generated realistic enhanced SDOCT data that is indistinguishable from SSOCT images providing improved visualization of the choroid. This technology enabled accurate measurements of choroidal metrics previously limited by the imaging depth constraints of SDOCT. The findings open new possibilities for utilizing affordable SDOCT devices in studying the choroid in both healthy and pathological conditions. [ABSTRACT FROM AUTHOR]

Accurate physical activity monitoring is essential to understand the impact of physical activity on one's physical health and overall well-being. However, advances in human activity recognition algorithms have been constrained by the limited availability of large labelled datasets. This study aims to leverage recent advances in self-supervised learning to exploit the large-scale UK Biobank accelerometer dataset—a 700,000 person-days unlabelled dataset—in order to build models with vastly improved generalisability and accuracy. Our resulting models consistently outperform strong baselines across eight benchmark datasets, with an F1 relative improvement of 2.5–130.9% (median 24.4%). More importantly, in contrast to previous reports, our results generalise across external datasets, cohorts, living environments, and sensor devices. Our open-sourced pre-trained models will be valuable in domains with limited labelled data or where good sampling coverage (across devices, populations, and activities) is hard to achieve. [ABSTRACT FROM AUTHOR]

The monitoring of physiological parameters is a crucial topic in promoting human health and an indispensable approach for assessing physiological status and diagnosing diseases. Particularly, it holds significant value for patients who require long-term monitoring or with underlying cardiovascular disease. To this end, Visual Contactless Physiological Monitoring (VCPM) is capable of using videos recorded by a consumer camera to monitor blood volume pulse (BVP) signal, heart rate (HR), respiratory rate (RR), oxygen saturation (SpO2) and blood pressure (BP). Recently, deep learning-based pipelines have attracted numerous scholars and achieved unprecedented development. Although VCPM is still an emerging digital medical technology and presents many challenges and opportunities, it has the potential to revolutionize clinical medicine, digital health, telemedicine as well as other areas. The VCPM technology presents a viable solution that can be integrated into these systems for measuring vital parameters during video consultation, owing to its merits of contactless measurement, cost-effectiveness, user-friendly passive monitoring and the sole requirement of an off-the-shelf camera. In fact, the studies of VCPM technologies have been rocketing recently, particularly AI-based approaches, but few are employed in clinical settings. Here we provide a comprehensive overview of the applications, challenges, and prospects of VCPM from the perspective of clinical settings and AI technologies for the first time. The thorough exploration and analysis of clinical scenarios will provide profound guidance for the research and development of VCPM technologies in clinical settings. [ABSTRACT FROM AUTHOR]

Big data serves as the cornerstone for constructing real-world deep learning systems across various domains. In medicine and healthcare, a single clinical site lacks sufficient data, thus necessitating the involvement of multiple sites. Unfortunately, concerns regarding data security and privacy hinder the sharing and reuse of data across sites. Existing approaches to multi-site clinical learning heavily depend on the security of the network firewall and system implementation. To address this issue, we propose Relay Learning, a secure deep-learning framework that physically isolates clinical data from external intruders while still leveraging the benefits of multi-site big data. We demonstrate the efficacy of Relay Learning in three medical tasks of different diseases and anatomical structures, including structure segmentation of retina fundus, mediastinum tumors diagnosis, and brain midline localization. We evaluate Relay Learning by comparing its performance to alternative solutions through multi-site validation and external validation. Incorporating a total of 41,038 medical images from 21 medical hosts, including 7 external hosts, with non-uniform distributions, we observe significant performance improvements with Relay Learning across all three tasks. Specifically, it achieves an average performance increase of 44.4%, 24.2%, and 36.7% for retinal fundus segmentation, mediastinum tumor diagnosis, and brain midline localization, respectively. Remarkably, Relay Learning even outperforms central learning on external test sets. In the meanwhile, Relay Learning keeps data sovereignty locally without cross-site network connections. We anticipate that Relay Learning will revolutionize clinical multi-site collaboration and reshape the landscape of healthcare in the future. [ABSTRACT FROM AUTHOR]

Congenital malformations of the central nervous system are among the most common major congenital malformations. Deep learning systems have come to the fore in prenatal diagnosis of congenital malformation, but the impact of deep learning-assisted detection of congenital intracranial malformations from fetal neurosonographic images has not been evaluated. Here we report a three-way crossover, randomized control trial (Trial Registration: ChiCTR2100048233) that assesses the efficacy of a deep learning system, the Prenatal Ultrasound Diagnosis Artificial Intelligence Conduct System (PAICS), in assisting fetal intracranial malformation detection. A total of 709 fetal neurosonographic images/videos are read interactively by 36 sonologists of different expertise levels in three reading modes: unassisted mode (without PAICS assistance), concurrent mode (using PAICS at the beginning of the assessment) and second mode (using PAICS after a fully unaided interpretation). Aided by PAICS, the average accuracy of the unassisted mode (73%) is increased by the concurrent mode (80%; P < 0.001) and the second mode (82%; P < 0.001). Correspondingly, the AUC is increased from 0.85 to 0.89 and to 0.90, respectively (P < 0.001 for all). The median read time per data is slightly increased in concurrent mode but substantially prolonged in the second mode, from 6 s to 7 s and to 11 s (P < 0.001 for all). In conclusion, PAICS in both concurrent and second modes has the potential to improve sonologists' performance in detecting fetal intracranial malformations from neurosonographic data. PAICS is more efficient when used concurrently for all readers. [ABSTRACT FROM AUTHOR]

Skin diseases affect one-third of the global population, posing a major healthcare burden. Deep learning may optimise healthcare workflows through processing skin images via neural networks to make predictions. A focus of deep learning research is skin lesion triage to detect cancer, but this may not translate to the wider scope of >2000 other skin diseases. We searched for studies applying deep learning to skin images, excluding benign/malignant lesions (1/1/2000-23/6/2022, PROSPERO CRD42022309935). The primary outcome was accuracy of deep learning algorithms in disease diagnosis or severity assessment. We modified QUADAS-2 for quality assessment. Of 13,857 references identified, 64 were included. The most studied diseases were acne, psoriasis, eczema, rosacea, vitiligo, urticaria. Deep learning algorithms had high specificity and variable sensitivity in diagnosing these conditions. Accuracy of algorithms in diagnosing acne (median 94%, IQR 86–98; n = 11), rosacea (94%, 90–97; n = 4), eczema (93%, 90–99; n = 9) and psoriasis (89%, 78–92; n = 8) was high. Accuracy for grading severity was highest for psoriasis (range 93–100%, n = 2), eczema (88%, n = 1), and acne (67–86%, n = 4). However, 59 (92%) studies had high risk-of-bias judgements and 62 (97%) had high-level applicability concerns. Only 12 (19%) reported participant ethnicity/skin type. Twenty-four (37.5%) evaluated the algorithm in an independent dataset, clinical setting or prospectively. These data indicate potential of deep learning image analysis in diagnosing and monitoring common skin diseases. Current research has important methodological/reporting limitations. Real-world, prospectively-acquired image datasets with external validation/testing will advance deep learning beyond the current experimental phase towards clinically-useful tools to mitigate rising health and cost impacts of skin disease. [ABSTRACT FROM AUTHOR]

Ileocolic intussusception is one of the common acute abdomens in children and is first diagnosed urgently using ultrasound. Manual diagnosis requires extensive experience and skill, and identifying surgical indications in assessing the disease severity is more challenging. We aimed to develop a real-time lesion visualization deep-learning pipeline to solve this problem. This multicenter retrospective-prospective study used 14,085 images in 8736 consecutive patients (median age, eight months) with ileocolic intussusception who underwent ultrasound at six hospitals to train, validate, and test the deep-learning pipeline. Subsequently, the algorithm was validated in an internal image test set and an external video dataset. Furthermore, the performances of junior, intermediate, senior, and junior sonographers with AI-assistance were prospectively compared in 242 volunteers using the DeLong test. This tool recognized 1,086 images with three ileocolic intussusception signs with an average of the area under the receiver operating characteristic curve (average-AUC) of 0.972. It diagnosed 184 patients with no intussusception, nonsurgical intussusception, and surgical intussusception in 184 ultrasound videos with an average-AUC of 0.956. In the prospective pilot study using 242 volunteers, junior sonographers' performances were significantly improved with AI-assistance (average-AUC: 0.966 vs. 0.857, P < 0.001; median scanning-time: 9.46 min vs. 3.66 min, P < 0.001), which were comparable to those of senior sonographers (average-AUC: 0.966 vs. 0.973, P = 0.600). Thus, here, we report that the deep-learning pipeline that guides lesions in real-time and is interpretable during ultrasound scanning could assist sonographers in improving the accuracy and efficiency of diagnosing intussusception and identifying surgical indications. [ABSTRACT FROM AUTHOR]

Preterm birth (PTB) is the leading cause of infant mortality globally. Research has focused on developing predictive models for PTB without prioritizing cost-effective interventions. Physical activity and sleep present unique opportunities for interventions in low- and middle-income populations (LMICs). However, objective measurement of physical activity and sleep remains challenging and self-reported metrics suffer from low-resolution and accuracy. In this study, we use physical activity data collected using a wearable device comprising over 181,944 h of data across N = 1083 patients. Using a new state-of-the art deep learning time-series classification architecture, we develop a 'clock' of healthy dynamics during pregnancy by using gestational age (GA) as a surrogate for progression of pregnancy. We also develop novel interpretability algorithms that integrate unsupervised clustering, model error analysis, feature attribution, and automated actigraphy analysis, allowing for model interpretation with respect to sleep, activity, and clinical variables. Our model performs significantly better than 7 other machine learning and AI methods for modeling the progression of pregnancy. We found that deviations from a normal 'clock' of physical activity and sleep changes during pregnancy are strongly associated with pregnancy outcomes. When our model underestimates GA, there are 0.52 fewer preterm births than expected (P = 1.01e − 67, permutation test) and when our model overestimates GA, there are 1.44 times (P = 2.82e − 39, permutation test) more preterm births than expected. Model error is negatively correlated with interdaily stability (P = 0.043, Spearman's), indicating that our model assigns a more advanced GA when an individual's daily rhythms are less precise. Supporting this, our model attributes higher importance to sleep periods in predicting higher-than-actual GA, relative to lower-than-actual GA (P = 1.01e − 21, Mann-Whitney U). Combining prediction and interpretability allows us to signal when activity behaviors alter the likelihood of preterm birth and advocates for the development of clinical decision support through passive monitoring and exercise habit and sleep recommendations, which can be easily implemented in LMICs. [ABSTRACT FROM AUTHOR]

The electrocardiogram (ECG) is the most frequently performed cardiovascular diagnostic test, but it is unclear how much information resting ECGs contain about long term cardiovascular risk. Here we report that a deep convolutional neural network can accurately predict the long-term risk of cardiovascular mortality and disease based on a resting ECG alone. Using a large dataset of resting 12-lead ECGs collected at Stanford University Medical Center, we developed SEER, the Stanford Estimator of Electrocardiogram Risk. SEER predicts 5-year cardiovascular mortality with an area under the receiver operator characteristic curve (AUC) of 0.83 in a held-out test set at Stanford, and with AUCs of 0.78 and 0.83 respectively when independently evaluated at Cedars-Sinai Medical Center and Columbia University Irving Medical Center. SEER predicts 5-year atherosclerotic disease (ASCVD) with an AUC of 0.67, similar to the Pooled Cohort Equations for ASCVD Risk, while being only modestly correlated. When used in conjunction with the Pooled Cohort Equations, SEER accurately reclassified 16% of patients from low to moderate risk, uncovering a group with an actual average 9.9% 10-year ASCVD risk who would not have otherwise been indicated for statin therapy. SEER can also predict several other cardiovascular conditions such as heart failure and atrial fibrillation. Using only lead I of the ECG it predicts 5-year cardiovascular mortality with an AUC of 0.80. SEER, used alongside the Pooled Cohort Equations and other risk tools, can substantially improve cardiovascular risk stratification and aid in medical decision making. [ABSTRACT FROM AUTHOR]

For hemodialysis patients, arteriovenous fistula (AVF) patency determines whether adequate hemofiltration can be achieved, and directly influences clinical outcomes. Here, we report the development and performance of a deep learning model for automated AVF stenosis screening based on the sound of AVF blood flow using supervised learning with data validated by ultrasound. We demonstrate the importance of contextualizing the sound with location metadata as the characteristics of the blood flow sound varies significantly along the AVF. We found the best model to be a vision transformer trained on spectrogram images. Our model can screen for stenosis at a performance level comparable to that of a nephrologist performing a physical exam, but with the advantage of being automated and scalable. In a high-volume, resource-limited clinical setting, automated AVF stenosis screening can help ensure patient safety via early detection of at-risk vascular access, streamline the dialysis workflow, and serve as a patient-facing tool to allow for at-home, self-screening. [ABSTRACT FROM AUTHOR]

Electronic health records are often incomplete, reducing the power of genetic association studies. For some diseases, such as knee osteoarthritis where the routine course of diagnosis involves an X-ray, image-based phenotyping offers an alternate and unbiased way to ascertain disease cases. We investigated this by training a deep-learning model to ascertain knee osteoarthritis cases from knee DXA scans that achieved clinician-level performance. Using our model, we identified 1931 (178%) more cases than currently diagnosed in the health record. Individuals diagnosed as cases by our model had higher rates of self-reported knee pain, for longer durations and with increased severity compared to control individuals. We trained another deep-learning model to measure the knee joint space width, a quantitative phenotype linked to knee osteoarthritis severity. In performing genetic association analysis, we found that use of a quantitative measure improved the number of genome-wide significant loci we discovered by an order of magnitude compared with our binary model of cases and controls despite the two phenotypes being highly genetically correlated. In addition we discovered associations between our quantitative measure of knee osteoarthritis and increased risk of adult fractures- a leading cause of injury-related death in older individuals-, illustrating the capability of image-based phenotyping to reveal epidemiological associations not captured in the electronic health record. For diseases with radiographic diagnosis, our results demonstrate the potential for using deep learning to phenotype at biobank scale, improving power for both genetic and epidemiological association analysis. [ABSTRACT FROM AUTHOR]

Duct-dependent congenital heart diseases (CHDs) are a serious form of CHD with a low detection rate, especially in underdeveloped countries and areas. Although existing studies have developed models for fetal heart structure identification, there is a lack of comprehensive evaluation of the long axis of the aorta. In this study, a total of 6698 images and 48 videos are collected to develop and test a two-stage deep transfer learning model named DDCHD-DenseNet for screening critical duct-dependent CHDs. The model achieves a sensitivity of 0.973, 0.843, 0.769, and 0.759, and a specificity of 0.985, 0.967, 0.956, and 0.759, respectively, on the four multicenter test sets. It is expected to be employed as a potential automatic screening tool for hierarchical care and computer-aided diagnosis. Our two-stage strategy effectively improves the robustness of the model and can be extended to screen for other fetal heart development defects. [ABSTRACT FROM AUTHOR]

Lung cancer screening using computed tomography (CT) has increased the detection rate of small pulmonary nodules and early-stage lung adenocarcinoma. It would be clinically meaningful to accurate assessment of the nodule histology by CT scans with advanced deep learning algorithms. However, recent studies mainly focus on predicting benign and malignant nodules, lacking of model for the risk stratification of invasive adenocarcinoma. We propose an ensemble multi-view 3D convolutional neural network (EMV-3D-CNN) model to study the risk stratification of lung adenocarcinoma. We include 1075 lung nodules (≤30 mm and ≥4 mm) with preoperative thin-section CT scans and definite pathology confirmed by surgery. Our model achieves a state-of-art performance of 91.3% and 92.9% AUC for diagnosis of benign/malignant and pre-invasive/invasive nodules, respectively. Importantly, our model outperforms senior doctors in risk stratification of invasive adenocarcinoma with 77.6% accuracy [i.e., Grades 1, 2, 3]). It provides detailed predictive histological information for the surgical management of pulmonary nodules. Finally, for user-friendly access, the proposed model is implemented as a web-based system (https://seeyourlung.com.cn). [ABSTRACT FROM AUTHOR]

Cerebrovascular disease is a leading cause of death globally. Prevention and early intervention are known to be the most effective forms of its management. Non-invasive imaging methods hold great promises for early stratification, but at present lack the sensitivity for personalized prognosis. Resting-state functional magnetic resonance imaging (rs-fMRI), a powerful tool previously used for mapping neural activity, is available in most hospitals. Here we show that rs-fMRI can be used to map cerebral hemodynamic function and delineate impairment. By exploiting time variations in breathing pattern during rs-fMRI, deep learning enables reproducible mapping of cerebrovascular reactivity (CVR) and bolus arrival time (BAT) of the human brain using resting-state CO2 fluctuations as a natural "contrast media". The deep-learning network is trained with CVR and BAT maps obtained with a reference method of CO2-inhalation MRI, which includes data from young and older healthy subjects and patients with Moyamoya disease and brain tumors. We demonstrate the performance of deep-learning cerebrovascular mapping in the detection of vascular abnormalities, evaluation of revascularization effects, and vascular alterations in normal aging. In addition, cerebrovascular maps obtained with the proposed method exhibit excellent reproducibility in both healthy volunteers and stroke patients. Deep-learning resting-state vascular imaging has the potential to become a useful tool in clinical cerebrovascular imaging. [ABSTRACT FROM AUTHOR]

The bold vision of AI-driven pervasive physiological monitoring, through the proliferation of off-the-shelf wearables that began a decade ago, has created immense opportunities to extract actionable information for precision medicine. These AI algorithms model input-output relationships of a system that, in many cases, exhibits complex nature and personalization requirements. A particular example is cuffless blood pressure estimation using wearable bioimpedance. However, these algorithms need training over significant amount of ground truth data. In the context of biomedical applications, collecting ground truth data, particularly at the personalized level is challenging, burdensome, and in some cases infeasible. Our objective is to establish physics-informed neural network (PINN) models for physiological time series data that would use minimal ground truth information to extract complex cardiovascular information. We achieve this by building Taylor's approximation for gradually changing known cardiovascular relationships between input and output (e.g., sensor measurements to blood pressure) and incorporating this approximation into our proposed neural network training. The effectiveness of the framework is demonstrated through a case study: continuous cuffless BP estimation from time series bioimpedance data. We show that by using PINNs over the state-of-the-art time series models tested on the same datasets, we retain high correlations (systolic: 0.90, diastolic: 0.89) and low error (systolic: 1.3 ± 7.6 mmHg, diastolic: 0.6 ± 6.4 mmHg) while reducing the amount of ground truth training data on average by a factor of 15. This could be helpful in developing future AI algorithms to help interpret pervasive physiologic data using minimal amount of training data. [ABSTRACT FROM AUTHOR]

The interpretation of lung auscultation is highly subjective and relies on non-specific nomenclature. Computer-aided analysis has the potential to better standardize and automate evaluation. We used 35.9 hours of auscultation audio from 572 pediatric outpatients to develop DeepBreath : a deep learning model identifying the audible signatures of acute respiratory illness in children. It comprises a convolutional neural network followed by a logistic regression classifier, aggregating estimates on recordings from eight thoracic sites into a single prediction at the patient-level. Patients were either healthy controls (29%) or had one of three acute respiratory illnesses (71%) including pneumonia, wheezing disorders (bronchitis/asthma), and bronchiolitis). To ensure objective estimates on model generalisability, DeepBreath is trained on patients from two countries (Switzerland, Brazil), and results are reported on an internal 5-fold cross-validation as well as externally validated (extval) on three other countries (Senegal, Cameroon, Morocco). DeepBreath differentiated healthy and pathological breathing with an Area Under the Receiver-Operator Characteristic (AUROC) of 0.93 (standard deviation [SD] ± 0.01 on internal validation). Similarly promising results were obtained for pneumonia (AUROC 0.75 ± 0.10), wheezing disorders (AUROC 0.91 ± 0.03), and bronchiolitis (AUROC 0.94 ± 0.02). Extval AUROCs were 0.89, 0.74, 0.74 and 0.87 respectively. All either matched or were significant improvements on a clinical baseline model using age and respiratory rate. Temporal attention showed clear alignment between model prediction and independently annotated respiratory cycles, providing evidence that DeepBreath extracts physiologically meaningful representations. DeepBreath provides a framework for interpretable deep learning to identify the objective audio signatures of respiratory pathology. [ABSTRACT FROM AUTHOR]

For the last 40 years, actigraphy or wearable accelerometry has provided an objective, low-burden and ecologically valid approach to assess real-world sleep and circadian patterns, contributing valuable data to epidemiological and clinical insights on sleep and sleep disorders. The proper use of wearable technology in sleep research requires validated algorithms that can derive sleep outcomes from the sensor data. Since the publication of the first automated scoring algorithm by Webster in 1982, a variety of sleep algorithms have been developed and contributed to sleep research, including many recent ones that leverage machine learning and / or deep learning approaches. However, it remains unclear how these algorithms compare to each other on the same data set and if these modern data science approaches improve the analytical validity of sleep outcomes based on wrist-worn acceleration data. This work provides a systematic evaluation across 8 state-of-the-art sleep algorithms on a common sleep data set with polysomnography (PSG) as ground truth. Despite the inclusion of recently published complex algorithms, simple regression-based and heuristic algorithms demonstrated slightly superior performance in sleep-wake classification and sleep outcome estimation. The performance of complex machine learning and deep learning models seem to suffer from poor generalization. This independent and systematic analytical validation of sleep algorithms provides key evidence on the use of wearable digital health technologies for sleep research and care. [ABSTRACT FROM AUTHOR]

Predicting recurrence in low-grade, early-stage endometrial cancer (EC) is both challenging and clinically relevant. We present a weakly-supervised deep learning framework, NaroNet, that can learn, without manual expert annotation, the complex tumor-immune interrelations at three levels: local phenotypes, cellular neighborhoods, and tissue areas. It uses multiplexed immunofluorescence for the simultaneous visualization and quantification of CD68 + macrophages, CD8 + T cells, FOXP3 + regulatory T cells, PD-L1/PD-1 protein expression, and tumor cells. We used 489 tumor cores from 250 patients to train a multilevel deep-learning model to predict tumor recurrence. Using a tenfold cross-validation strategy, our model achieved an area under the curve of 0.90 with a 95% confidence interval of 0.83–0.95. Our model predictions resulted in concordance for 96,8% of cases (κ = 0.88). This method could accurately assess the risk of recurrence in EC, outperforming current prognostic factors, including molecular subtyping. [ABSTRACT FROM AUTHOR]

AASM guidelines are the result of decades of efforts aiming at standardizing sleep scoring procedure, with the final goal of sharing a worldwide common methodology. The guidelines cover several aspects from the technical/digital specifications, e.g., recommended EEG derivations, to detailed sleep scoring rules accordingly to age. Automated sleep scoring systems have always largely exploited the standards as fundamental guidelines. In this context, deep learning has demonstrated better performance compared to classical machine learning. Our present work shows that a deep learning-based sleep scoring algorithm may not need to fully exploit the clinical knowledge or to strictly adhere to the AASM guidelines. Specifically, we demonstrate that U-Sleep, a state-of-the-art sleep scoring algorithm, can be strong enough to solve the scoring task even using clinically non-recommended or non-conventional derivations, and with no need to exploit information about the chronological age of the subjects. We finally strengthen a well-known finding that using data from multiple data centers always results in a better performing model compared with training on a single cohort. Indeed, we show that this latter statement is still valid even by increasing the size and the heterogeneity of the single data cohort. In all our experiments we used 28528 polysomnography studies from 13 different clinical studies. [ABSTRACT FROM AUTHOR]

Deep Reinforcement Learning (DRL) has been increasingly attempted in assisting clinicians for real-time treatment of sepsis. While a value function quantifies the performance of policies in such decision-making processes, most value-based DRL algorithms cannot evaluate the target value function precisely and are not as safe as clinical experts. In this study, we propose a Weighted Dueling Double Deep Q-Network with embedded human Expertise (WD3QNE). A target Q value function with adaptive dynamic weight is designed to improve the estimate accuracy and human expertise in decision-making is leveraged. In addition, the random forest algorithm is employed for feature selection to improve model interpretability. We test our algorithm against state-of-the-art value function methods in terms of expected return, survival rate, action distribution and external validation. The results demonstrate that WD3QNE obtains the highest survival rate of 97.81% in MIMIC-III dataset. Our proposed method is capable of providing reliable treatment decisions with embedded clinician expertise. [ABSTRACT FROM AUTHOR]

The increasing availability of large institutional and public histopathology image datasets is enabling the searching of these datasets for diagnosis, research, and education. Although these datasets typically have associated metadata such as diagnosis or clinical notes, even carefully curated datasets rarely contain annotations of the location of regions of interest on each image. As pathology images are extremely large (up to 100,000 pixels in each dimension), further laborious visual search of each image may be needed to find the feature of interest. In this paper, we introduce a deep-learning-based reverse image search tool for histopathology images: Similar Medical Images Like Yours (SMILY). We assessed SMILY's ability to retrieve search results in two ways: using pathologist-provided annotations, and via prospective studies where pathologists evaluated the quality of SMILY search results. As a negative control in the second evaluation, pathologists were blinded to whether search results were retrieved by SMILY or randomly. In both types of assessments, SMILY was able to retrieve search results with similar histologic features, organ site, and prostate cancer Gleason grade compared with the original query. SMILY may be a useful general-purpose tool in the pathologist's arsenal, to improve the efficiency of searching large archives of histopathology images, without the need to develop and implement specific tools for each application. [ABSTRACT FROM AUTHOR]

In recent years, there is increasing enthusiasm in the healthcare research community for artificial intelligence to provide big data analytics and augment decision making. One of the prime reasons for this is the enormous impact of deep learning for utilization of complex healthcare big data. Although deep learning is a powerful analytic tool for the complex data contained in electronic health records (EHRs), there are also limitations which can make the choice of deep learning inferior in some healthcare applications. In this paper, we give a brief overview of the limitations of deep learning illustrated through case studies done over the years aiming to promote the consideration of alternative analytic strategies for healthcare. [ABSTRACT FROM AUTHOR]