Search

The adoption of electronic health records (EHRs) has made patient data increasingly accessible, precipitating the development of various clinical decision support systems and data-driven models to help physicians. However, missing data are common in EHR-derived datasets, which can introduce significant uncertainty, if not invalidating the use of a predictive model. Machine learning (ML)-based imputation methods have shown promise in various domains for the task of estimating values and reducing uncertainty to the point that a predictive model can be employed. We introduce Autopopulus, a novel framework that enables the design and evaluation of various autoencoder architectures for efficient imputation on large datasets. Autopopulus implements existing autoencoder methods as well as a new technique that outputs a range of estimated values (rather than point estimates), and demonstrates a workflow that helps users make an informed decision on an appropriate imputation method. To further illustrate Autopopulus' utility, we use it to identify not only which imputation methods can most accurately impute on a large clinical dataset, but to also identify the imputation methods that enable downstream predictive models to achieve the best performance for prediction of chronic kidney disease (CKD) progression.

Cancer screening can benefit from individualized decision-making tools that decrease overdiagnosis. The heterogeneity of cancer screening participants advocates the need for more personalized methods. Partially observable Markov decision processes (POMDPs), when defined with an appropriate reward function, can be used to suggest optimal, individualized screening policies. However, determining an appropriate reward function can be challenging. Here, we propose the use of inverse reinforcement learning (IRL) to form rewards functions for lung and breast cancer screening POMDPs. Using experts (physicians) retrospective screening decisions for lung and breast cancer screening, we developed two POMDP models with corresponding reward functions. Specifically, the maximum entropy (MaxEnt) IRL algorithm with an adaptive step size was employed to learn rewards more efficiently; and combined with a multiplicative model to learn state-action pair rewards for a POMDP. The POMDP screening models were evaluated based on their ability to recommend appropriate screening decisions before the diagnosis of cancer. The reward functions learned with the MaxEnt IRL algorithm, when combined with POMDP models in lung and breast cancer screening, demonstrate performance comparable to experts. The Cohen’s Kappa score of agreement between the POMDPs and physicians’ predictions was high in breast cancer and had a decreasing trend in lung cancer.

Risk prediction models are crucial for assessing the pretest probability of cancer and are applied to stratify patient management strategies. These models are frequently based on multivariate regression analysis, requiring that all risk factors be specified, and do not convey the confidence in their predictions. We present a framework for uncertainty analysis that accounts for variability in input values. Uncertain or missing values are replaced with a range of plausible values. These ranges are used to compute individualized risk confidence intervals. We demonstrate our approach using the Gail model to evaluate the impact of uncertainty on management decisions. Up to 13% of cases (uncertain) had a risk interval that falls within the decision threshold (e.g., 1.67% 5-year absolute risk). A small number of cases changed from low- to high-risk when missing values were present. Our analysis underscores the need for better communication of input assumptions that influence the resulting predictions.

A growing number of individuals who are considered at high risk of cancer are now routinely undergoing population screening. However, noted harms such as radiation exposure, overdiagnosis, and overtreatment underscore the need for better temporal models that predict who should be screened and at what frequency. The mean sojourn time (MST), an average duration period when a tumor can be detected by imaging but with no observable clinical symptoms, is a critical variable for formulating screening policy. Estimation of MST has been long studied using continuous Markov model (CMM) with Maximum likelihood estimation (MLE). However, a lot of traditional methods assume no observation error of the imaging data, which is unlikely and can bias the estimation of the MST. In addition, the MLE may not be stably estimated when data is sparse. Addressing these shortcomings, we present a probabilistic modeling approach for periodic cancer screening data. We first model the cancer state transition using a three state CMM model, while simultaneously considering observation error. We then jointly estimate the MST and observation error within a Bayesian framework. We also consider the inclusion of covariates to estimate individualized rates of disease progression. Our approach is demonstrated on participants who underwent chest x-ray screening in the National Lung Screening Trial (NLST) and validated using posterior predictive p-values and Pearson's chi-square test. Our model demonstrates more accurate and sensible estimates of MST in comparison to MLE.

A growing number of individuals who are considered at high risk of cancer are now routinely undergoing population screening. However, noted harms such as radiation exposure, overdiagnosis, and overtreatment underscore the need for better temporal models that predict who should be screened and at what frequency. The mean sojourn time (MST), an average duration period when a tumor can be detected by imaging but with no observable clinical symptoms, is a critical variable for formulating screening policy. Estimation of MST has been long studied using continuous Markov model (CMM) with Maximum likelihood estimation (MLE). However, a lot of traditional methods assume no observation error of the imaging data, which is unlikely and can bias the estimation of the MST. In addition, the MLE may not be stably estimated when data is sparse. Addressing these shortcomings, we present a probabilistic modeling approach for periodic cancer screening data. We first model the cancer state transition using a three state CMM model, while simultaneously considering observation error. We then jointly estimate the MST and observation error within a Bayesian framework. We also consider the inclusion of covariates to estimate individualized rates of disease progression. Our approach is demonstrated on participants who underwent chest x-ray screening in the National Lung Screening Trial (NLST) and validated using posterior predictive p-values and Pearson’s chi-square test. Our model demonstrates more accurate and sensible estimates of MST in comparison to MLE.

IntroductionIdentifying high-risk lung cancer individuals at an early disease stage is the most effective way of improving survival. The landmark National Lung Screening Trial (NLST) demonstrated the utility of low-dose computed tomography (LDCT) imaging to reduce mortality (relative to x-ray screening). As a result of the NLST and other studies, imaging-based lung cancer screening programs are now being implemented. However, LDCT interpretation results in a high number of false positives. A set of dynamic Bayesian networks (DBN) were designed and evaluated to provide insight into how longitudinal data can be used to help inform lung cancer screening decisions.MethodsThe LDCT arm of the NLST dataset was used to build and explore five DBNs for high-risk individuals. Three of these DBNs were built using a backward construction process, and two using structure learning methods. All models employ demographic, smoking status, cancer history, family lung cancer history, exposure risk factors, comorbidities related to lung cancer, and LDCT screening outcome information. Given the uncertainty arising from lung cancer screening, a cancer state-space model based on lung cancer staging was utilized to characterize the cancer status of an individual over time. The models were evaluated on balanced training and test sets of cancer and non-cancer cases to deal with data imbalance and overfitting.ResultsResults were comparable to expert decisions. The average area under the curve (AUC) of the receiver operating characteristic (ROC) for the three intervention points of the NLST trial was higher than 0.75 for all models. Evaluation of the models on the complete LDCT arm of the NLST dataset (N = 25, 486) demonstrated satisfactory generalization. Consensus of predictions over similar cases is reported in concordance statistics between the models’ and the physicians’ predictions. The models’ predictive ability with respect to missing data was also evaluated with the sample of cases that missed the second screening exam of the trial (N = 417). The DBNs outperformed comparison models such as logistic regression and naïve Bayes.ConclusionThe lung cancer screening DBNs demonstrated high discrimination and predictive power with the majority of cancer and non-cancer cases.

Continuous renal replacement therapy (CRRT) is a form of dialysis prescribed to severely ill patients who cannot tolerate regular hemodialysis. However, as the patients are typically very ill to begin with, there is always uncertainty as to whether they will survive during or after CRRT treatment. Because of outcome uncertainty, a large percentage of patients treated with CRRT do not survive, utilizing scarce resources and raising false hope in patients and their families. To address these issues, we present a machine-learning-based algorithm to predict if patients will survive after being treated with CRRT. We use information extracted from electronic health records from patients who were placed on CRRT at multiple institutions to train a model that predicts CRRT survival outcome; on a held-out test set, the model achieved an area under the receiver operating curve of 0.929 (CI=0.917-0.942). Feature importance, error, and subgroup analyses identified consistently, mean corpuscular volume as a driving feature for model predictions. Overall, we demonstrate the potential for predictive machine-learning models to assist clinicians in alleviating the uncertainty of CRRT patient survival outcomes, with opportunities for future improvement through further data collection and advanced modeling., Competing Interests: Additional Declarations: There is NO Competing Interest.

The adoption of electronic health records (EHRs) has made patient data increasingly accessible, precipitating the development of various clinical decision support systems and data-driven models to help physicians. However, missing data are common in EHR-derived datasets, which can introduce significant uncertainty, if not invalidating the use of a predictive model. Machine learning (ML)-based imputation methods have shown promise in various domains for the task of estimating values and reducing uncertainty to the point that a predictive model can be employed. We introduce Autopopulus, a novel framework that enables the design and evaluation of various autoencoder architectures for efficient imputation on large datasets. Autopopulus implements existing autoencoder methods as well as a new technique that outputs a range of estimated values (rather than point estimates), and demonstrates a workflow that helps users make an informed decision on an appropriate imputation method. To further illustrate Autopopulus' utility, we use it to identify not only which imputation methods can most accurately impute on a large clinical dataset, but to also identify the imputation methods that enable downstream predictive models to achieve the best performance for prediction of chronic kidney disease (CKD) progression.

Risk prediction models are crucial for assessing the pretest probability of cancer and are applied to stratify patient management strategies. These models are frequently based on multivariate regression analysis, requiring that all risk factors be specified, and do not convey the confidence in their predictions. We present a framework for uncertainty analysis that accounts for variability in input values. Uncertain or missing values are replaced with a range of plausible values. These ranges are used to compute individualized risk confidence intervals. We demonstrate our approach using the Gail model to evaluate the impact of uncertainty on management decisions. Up to 13% of cases (uncertain) had a risk interval that falls within the decision threshold (e.g., 1.67% 5-year absolute risk). A small number of cases changed from low- to high-risk when missing values were present. Our analysis underscores the need for better communication of input assumptions that influence the resulting predictions.