Machine learning for predicting readmission risk among the frail: Explainable AI for healthcare

Abstract

Highlights d We report a 0.79 AUROC for 30-day readmission prediction d We use frailty, comorbidity, high-risk medications, and demographics for improved accuracy d We identify clusters of high-risk patients based on sets of patient features d Explainability is a prime focus for model predictions at different levels of granularity

Frailty in Aging and the Search for the Optimal Biomarker: A Review

“…AI can be defined as a branch of computer science whose objective is to create systems or methods that analyze information and allow the management of complexity in a wide range of applications [ 188 ]. Previous studies suggest that the use of a predictive model based on machine learning could be useful to detect future frailty conditions, as well as the risk of hospital readmissions of these patients, using both clinical and socioeconomic variables that can generally be collected in centers for health care [ 189 , 190 ].…”

Machine Learning Approaches for the Frailty Screening: A Narrative Review

“…In addition, a score or set of criteria was used, developed, and validated to identify frailty. The most common frailty instruments used in research and clinical practice are the Fried frailty phenotype (FP), which is based on five items (slow walking speed, weak grip strength, low physical activity, unintended weight loss, and exhaustion), minimum of three of five criteria for classifying as frailty [ 21 , 22 , 23 ]. Nevertheless, there is insufficient evidence to determine the best tool for use in research and clinical practice [ 20 ].…”

“…Moreover, these measures require specialized equipment (e.g., dynamometer to grip strength), not always clinically viable (e.g., for patients with dementia), and also require a manual evaluation process (e.g., timed-get-up-and-go) that is subject to operator error due to the need for training beyond time to administer [ 25 ]. Furthermore, the prevalence of frailty varies across settings and adopted tests, making it difficult to scale to the population level [ 21 , 22 , 25 , 26 ]. In this view, an alternative is exploring approaches to screening frailty from routinely collected data (e.g., medical claims, prescriptions, administrative data, and individual records) [ 25 ].…”

“…Such decision trees are designed to have a low correlation to each other to encourage range among the trees. Moreover, RFs use the rules of bootstrapping and aggregating to build trees based on several subsets of the training data using different subsets of features [ 21 , 35 ].…”

“…These models use boosting, an ensemble method where each tree or model corrects errors made by earlier trees. XGBoost requires minimal feature engineering, allowing steps such as normalizations and scaling to be omitted, and outliers have little impact [ 21 , 35 , 36 ].…”