Search

• A novel model predicts early antibody-incompatible kidney transplant rejection. • The models were trained on a small dataset of pre-transplant characteristics. • Decision Tree and Random Forest classifiers achieved 85% accuracy. • The models identified key risk factors, including specific IgG subclass levels. Clinical datasets are commonly limited in size, thus restraining applications of Machine Learning (ML) techniques for predictive modelling in clinical research and organ transplantation. We explored the potential of Decision Tree (DT) and Random Forest (RF) classification models, in the context of small dataset of 80 samples, for outcome prediction in high-risk kidney transplantation. The DT and RF models identified the key risk factors associated with acute rejection: the levels of the donor specific IgG antibodies, the levels of IgG4 subclass and the number of human leucocyte antigen mismatches between the donor and recipient. Furthermore, the DT model determined dangerous levels of donor-specific IgG subclass antibodies, thus demonstrating the potential of discovering new properties in the data when traditional statistical tools are unable to capture them. The DT and RF classifiers developed in this work predicted early transplant rejection with accuracy of 85%, thus offering an accurate decision support tool for doctors tasked with predicting outcomes of kidney transplantation in advance of the clinical intervention. [ABSTRACT FROM AUTHOR]

Advancements in health technologies are crucial to support healthcare professionals, improve patient outcomes, and best utilize increasingly scarce and under-demand healthcare resources. This research presents an initial proof-of-concept study of simple, non-invasive monitoring techniques used in Mechanical Ventilation (MV), which is the primary therapy for Acute Respiratory Distress Syndrome (ARDS). The high levels of inter-patient variability seen in patients with ARDS have resulted in much speculation about the ideal method of determining ventilation settings, such as tidal volume (Vt) and Positive End Expiratory Pressure (PEEP). One of the oldest and simplest methods is acoustic sensing of recruitment and lung condition. This project involves using a digital recording stethoscope to monitor the acoustic output of patients in the Intensive Care Unit (ICU) during mechanical lung ventilation. During lung recruitment, ‘crackles’ can be heard within the chest cavity with a stethoscope. These crackles vary significantly, depending on the status of the patient’s respiratory system and are used as an indicator of the level of alveolar recruitment. This preliminary, proof-of-concept study focused on crackle detection and involved gathering sound samples from patients in the Christchurch Hospital ICU with evidence of crackles in the chest cavity. Frequency based analysis showed that crackles can be detected as emissions with higher power levels between 100 and 300 Hz (subject to patient variability). The ability to non-invasively record, detect and quantify the intensity of crackles could provide immediate feedback to clinicians and, in the long term, aid in the optimization of ventilator therapy. [ABSTRACT FROM AUTHOR]

Highlights: [•] 14 virtual patients were created simulating intra-patient variability. [•] Blood glucose (BG) and continuous monitor (CGM) data were simulated. [•] Identification using interval models for consideration of uncertainty was performed. [•] Linear regressions between CGM and BG widths were calculated for width estimation. [•] Prediction capabilities of the models with the estimated widths were evaluated. [ABSTRACT FROM AUTHOR]

Background: Minimizing excessive postprandial glucose excursion is still an unmet need in treating diabetes. This work addresses the analysis and modeling in the postprandial state of two commercial CGM devices. Methods: Twelve patients with type 1 diabetes were studied in the postprandial state on four different occasions under controlled conditions. Each time, we performed simultaneous glucose monitoring using the Dexcom SEVEN® PLUS (47 datasets) and the Medtronic Paradigm® Veo™ (42 datasets). The following statistical properties of the error signal were analyzed and modeled sequentially for the two devices: the lag time, the error stationarity, the error probability distribution and the time correlation. Finally, models were built for sensor simulation in silico studies. Results: Lag time followed an exponential probability distribution for both monitors (μ sevenplus =108, μ veo =1.69). Standard deviation and mean of the error signal, calculated as time-dependent signals across the population of sensors, were time-varying and correlated with the reference value and its rate of change, respectively. In both cases, a high variability of postprandial behaviors was observed. After non-stationarity compensation, a logistic distribution was obtained for the SEVEN® PLUS error (close to normal distribution). Regarding the Paradigm® Veo™, a multimodal distribution was obtained, which turned into normal after elimination of five “unstable” sensors. Finally, a first order autoregressive model fitted the SEVEN® PLUS error time-series while a third-order filter was necessary for the Paradigm® Veo™. Conclusions: The Paradigm® Veo™ device exhibited greater delay variability with higher delay time and higher probability of abnormal sensor behaviors as compared to the SEVEN® PLUS device. In both cases, the observed variability may have important clinical implications in postprandial performance. Therefore, further improvements are needed in calibration algorithms to reduce this variability. [ABSTRACT FROM AUTHOR]

Abstract: The problem of determining an oral dose, or schedule of oral doses, that gives rise to an arbitrary area-under-curve or to points on the time-series for a variable of interest in a drug kinetics model is considered. These two measures are considered as surrogates for the particular drug response to the dose. The approach taken is to formulate the problem as a fixed point one to which a version of the contraction mapping theorem can be applied. The results, illustrated for a model for the anti-cancer agent topotecan, demonstrate the applicability of the approach. [ABSTRACT FROM AUTHOR]

Abstract: The control of blood pressure is a complex mixture of neural, hormonal and intrinsic interactions at the level of the heart, kidney and blood vessels. While experimental approaches to understanding these interactions are useful, it remains difficult to conduct experiments to quantify these interactions as the number of parameters increases. Thus, modelling of such physiological systems can offer considerable assistance. Typical mathematical models which describe the ability of the blood vessels to change their diameter (vasoconstriction) assume linearity of operation. However, due to the interaction of multiple vasocontrictive and vasodilative effectors, there is a significant nonlinear response to the influence of neural factors, particularly at higher levels of nerve activity (often seen in subjects with high blood pressure) which leads to low blood flow rates. This paper proposes a number of nonlinear mathematical models for the relationship between neural influences (sympathetic nerve activity (SNA)) and renal blood flow, using a feedback path to model the predominantly nonlinear effect of local vasoactive modulators such as nitric oxide, which oppose the action of SNA. The model structures are motivated by basic physiological principles, while the model parameters are determined using numerical optimisation techniques using open-loop data collected from rabbits. The models were verified by demonstrating correlation between experimental results and model outputs. [Copyright &y& Elsevier]