Using Explainable AI to understand frailty indicators

Abstract

Introduction: The prevalence of frailty is on the rise with the aging population and increasing life expectancy, which often is accompanied by comorbidities. Frailty can be effectively detected using Frailty index such as KCL index. Early detection of frailty allows applying measures that reduce the conversion rate to frail, and improve the quality of life in the frail people. Therefore, to facilitate the screening of frailty status at the primary care level, we suggest to produce a shorter version of the KCL questionnaire. Aim: To understand the importance of KCL components in the decision making process for frailty and use machine learning approach to shorten the Questionnaire while maintaining reasonable accuracy, making it easier to screen for frailty in primary care. Methods: We developed an automated framework of three steps: Feature importance determination using Shap values, testing models with Cross-validation with increased addition of selected features. Moreover, we validated the reliability of KCL to detect frailty by comparing the results of KCL criteria with the unsupervised clustering of the data. Results: Our approach allowed us to identify the most important questions in the KCL questionnaire and demonstrate its performance using a short version with only four questions (4) Do you visit homes of friends?, (6) Are you able to go upstairs without using handrails or the wall for support? (10) Do you feel anxious about falling when you walk?, and (25) (In the past two weeks) Have you felt exhausted for no apparent reason?). We also showed that the data clustering corresponds well with the results of KCL criteria. Discussion and Conclusion: While it is difficult to predict pre-frail status using shorter KCL questionnaire, it was shown to be fairly accurate in predicting frail status using only four questions.