Search

Heart failure (HF) is among the leading causes of death. Its prevalence is increasing dramatically causing considerable healthcare costs as well. Remote patient monitoring (RPM) is one of the solutions to enhance patient well-being. Taking advantage of new advances in artificial intelligence and RPM digital platforms, we propose HeartPredict: a novel machine learning algorithm for early detection of HF episodes and associated unplanned hospitalizations. The algorithm relies on a telemonitoring dataset from the largest European clinical study on HF. It uses balanced random forests on significant features extracted from patient weight time series, symptoms and socio-demographics. We benchmark HeartPredict with rules from medical guidelines and state of the art machine learning models. HeartPredict has better performance in terms of sensitivity (72% vs. guidelines: 53%) and specificity (94% vs. guidelines: 84%), with AUROC = 0.8. In addition, we introduce precocity as a new criterion to evaluate the ability of the algorithm to detect early HF risks, and therefore, enabling proactive medical actions. Finally, we show that HeartPredict prediction scores are consistent with HF risk levels, thereby limiting the risk of non-detection for patients.

Introduction: Recent advances in Machine Learning (M.L.) could be valuable to predict hospitalisations. Severe Chronic Obstructive Pulmonary Disease (COPD) is associated with high hospitalisation rate. Objective: To develop predictive models using M.L. approach and evaluate their accuracy using data from the COPD Patient Management European Trial (COMET, ClinicalTrials.gov NCT01241526). Methods: The first step of development focused on data pre-processing. The raw data sets included socio-demographics, comorbidities, acute care hospitalisation history, lab tests, prescribed medication, vital signs and symptoms questionnaires. The second step used an ensemble trees-based algorithm to identify the most important predictive variables. Using the subset of relevant variables, the third step implemented a predictive model based on a voting strategy to combine most performing M.L. algorithms results. Results: Data sets of 345 severe COPD patients were collected during their mean 12 months follow up leading to 367 acute care hospitalisations. The predictive model, evaluated for hospital admission in the next 7 days, achieved 78% of sensitivity (Se) and 76% of specificity (Sp). Performance of the model was also evaluated for the prediction of hospital readmissions, at 30 days (Se=65%; Sp=64%), 60 days (Se=68%; Sp=67%) and 90 days (Se=71%; Sp=71%) after discharge. Variables weighing the most in the model were time since previous hospitalisation, occurrence of exacerbations and the smoking status. Conclusions: M.L. algorithms could be powerful tools to predict the risk of hospitalisations in vulnerable patients. Training predictive models with limited size of data sets is the main difficulty encountered.

Introduction: Early detection of exacerbations in Chronic Obstructive Pulmonary Disease (COPD) is a research challenge. Objective: Evaluate a machine learning (M.L.) approach for detection of COPD exacerbations and identification of the most relevant predictors. Methods: Data from a subset of 173 patients on home oxygen included in the COPD Patient Management European Trial (COMET, ClinicalTrials. gov NCT01241526). The model includes socio-demographics, previous exacerbations, medication and records from the NOWOX oxygen monitoring device (respiratory rate (RR) and daily oxygen treatment duration (TD)). Data were processed using a 3 to 20 days slide window before exacerbations versus days without exacerbation. Comparison of the most recent M.L. approaches was performed using a split of the data into train-test-validation sets, allowing to develop a new predictive algorithm. Results: 425 moderate to severe exacerbations occurred during the mean 12 months follow-up. Our algorithm achieved 70% of sensitivity and 71% of specificity, in a time window ranging from 1 to 5 days prior the first day of the exacerbation. Increasing in RR and TD, time since last exacerbation and number of previous exacerbations were the most relevant parameters. Conclusions: Although this evaluation is performed using only easy-to-access data of a limited number of patients, the M.L. algorithm shows promising results. Providing additional data will increase performance.

In the context of evolutionary data classification, dynamical modeling techniques are useful to continuously learn clusters models. Dedicated to on-line clustering, the AUDyC (Auto-adaptive and Dynamical Clustering) algorithm is an unsupervised neural network with auto-adaptive abilities in nonstationary environment. These particular abilities are based on specific learning rules that are developed into three stages: “Classification”, “Evaluation” and “Fusion”. In this paper, we propose a new densities merge mechanism to improve the “Fusion” stage in order to avoid some local optima drawbacks of Gaussian fitting. The novelty of our approach is to use an ambiguity rule of fuzzy modelling with new merge acceptance criteria. Our approach can be generalized to any type of fuzzy classification method using Gaussian models. Some experiments are presented to show the efficiency of our approach to circumvent to AUDyC NN local optima problems.