Towards Clinical Prediction with Transparency: An Explainable AI Approach to Survival Modelling in Residential Aged Care

Abstract

AbstractBackgroundAn accurate estimate of expected survival time assists people near the end of life to make informed decisions about their medical care.ObjectivesUse advanced machine learning methods to develop an interpretable survival model for older people admitted to residential age care.SettingA large Australasian provider of residential age care services.ParticipantsAll residents aged 65 years and older, admitted for long-term residential care between July 2017 and August 2023.Sample size11,944 residents from 40 individual care facilities.PredictorsAge category, gender, predictors related to falls, health status, co-morbidities, cognitive function, mood state, nutritional status, mobility, smoking history, sleep, skin integrity, and continence.OutcomeProbability of survival at all time points post-admission. The final model is calibrated to estimate the probability of survival at 6 months post-admission.Statistical AnalysisCox Proportional Hazards (CoxPH), Elastic Net (EN), Ridge Regression (RR), Lasso, Gradient Boosting (GB), XGBoost (XGB) and Random Forest (RF) were tested in 20 experiments using different train/test splits at a 90/10 ratio. Model accuracy was evaluated with the Concordance Index (C-index), Harrell’s C-index, dynamic AUROC, Integrated Bier Score (IBS) and calibrated ROC analysis. XGBoost was selected as the optimal model and calibrated for time-specific predictions at 1,3,6 and 12 months post admission using Platt scaling. SHapley Additive exPlanations (SHAP) values from the 6-month model were plotted to demonstrate the global and local effect of specific predictors on survival probabilities.ResultsFor predicting survival across all time periods the GB, XGB and RF ensemble models had the best C-Index values of 0.714, 0.712 and 0.712 respectively. We selected the XGB model for further development and calibration and to provide interpretable outputs. The calibrated XGB model had a dynamic AUROC, when predicting survival at 6-months, of 0.746 (95% CI 0.744-0.749). For individuals with a 0.2 survival probability (80% risk of death within 6-months) the model had a negative predictive value of 0.74. Increased age, male gender, reduced mobility, poor general health status, elevated pressure ulcer risk, and lack of appetite were identified as the strongest predictors of imminent mortality.ConclusionsThis study demonstrates the effective application of machine learning in developing a survival model for people admitted to residential aged care. The model has adequate predictive accuracy and confirms clinical intuition about specific mortality risk factors at both the cohort and the individual level. Advancements in explainable AI, as demonstrated in this study, not only improve clinical usability of machine learning models by increasing transparency about how predictions are generated but may also reveal novel clinical insights.SUMMARY BOXSection 1: What is already known on this topicExisting models for estimating survival in aged care settings have been primarily based on prognostic indices which do not have advanced capabilities of machine learning approaches.There is a notable absence of both machine learning and AI tools that provide high interpretability of models and their predictions in residential aged care settings, crucial for clinical decision-making.Section 2: What this study addsOur study applies and demonstrates the utility of machine learning models for survival prediction in residential aged care settings, with a focus on the six month survival probabilities.The study performs extensive experiments using numerous algorithms, and demonstrates how multiple tools can be used in concert to provide personalized and highly interpretable predictions that enable clinicians to discuss care preferences with patients and families in an informed manner.This research sets a benchmark on how various AI technologies can be integrated with machine learning to offer effective solutions and greater transparency for clinical decision-making in aged care settings specifically, and predictive healthcare analytics more generally.