Heart Failure Diagnosis, Readmission, and Mortality Prediction Using Machine Learning and Artificial Intelligence Models

Abstract

Purpose of Review One in five people will develop heart failure (HF), and 50% of HF patients die in 5 years. The HF diagnosis, readmission, and mortality prediction are essential to develop personalized prevention and treatment plans. This review summarizes recent findings and approaches of machine learning models for HF diagnostic and outcome prediction using electronic health record (EHR) data. Recent Findings A set of machine learning models have been developed for HF diagnostic and outcome prediction using diverse variables derived from EHR data, including demographic, medical note, laboratory, and image data, and achieved expert-comparable prediction results. Summary Machine learning models can facilitate the identification of HF patients, as well as accurate patient-specific assessment of their risk for readmission and mortality. Additionally, novel machine learning techniques for integration of diverse data and improvement of model predictive accuracy in imbalanced data sets are critical for further development of these promising modeling methodologies.

An evidence-based data science perspective on the prediction of heart failure readmissions

“…While the variation in methods, data, and performance among HFR studies has been thoroughly reviewed by others 3,4,10 , few explanations have been given for the shared lack of success or the commonalities that have contributed to it. In the current work, we identify and discuss common factors contributing to the difficulty in predicting HFR, 30-day all-cause or otherwise.…”

“…Among the comparatively few studies that report it, precision typically ranges from 0.09 to 0.44, meaning that 56 – 91% of predictions for readmission are often incorrect. 3,21,24,28,29 If hospitals acted on the results of such predictions, they would invest their limited time and resources into expectations that are, more often than not, false alarms.…”

“…1 Consequently, to aid hospital enterprises in reducing HFR, hundreds of studies have used or developed indices and models to predict HFR, 30-day all-cause or otherwise. [2][3][4] These HFR studies have varied greatly in their readmission time frames (30d -1yr) and methodologies (risk indices, statistical models, machine learning), as well as in the data underpinning their predictions (clinical, administrative, psychosocial). [4][5][6][7][8] However, the degrees of performance and applicability needed to make transformative progress in predicting HFR have been notoriously difficult to achieve 4,9 , begging the question of why.…”

“…Surprisingly, many HFR studies never mention or measure precision. [2][3][4]7,8,[13][14][15][16]18,20,[23][24][25][26][27] However, without precision there is little way to establish confidence in whether or not a prediction for readmission can be trusted. Among the comparatively few studies that report it, precision typically ranges from 0.09 to 0.44, meaning that 56 -91% of predictions for readmission are often incorrect.…”

Image_1.pdf

“…Deep learning algorithms can effectively capture the informative and useful features and patterns from the rich healthcare information in EHR data (7). For example, a very recent study showed that deep-learning-based model achieved significantly higher accuracy to predict mortality among acute heart failure patients than the existing score models and several machine learning models by using EHR data (8)(9)(10)(11)(12)(13).…”