Use of multiprognostic index domain scores, clinical data, and machine learning to improve 12-Month mortality risk prediction in older hospitalized patients: Prospective cohort study

Background: The Multidimensional Prognostic Index (MPI) is an aggregate, comprehensive, geriatric assessment scoring system derived from eight domains that predict adverse outcomes, including 12-month mortality. However, the prediction accuracy of using the three MPI categories (mild, moderate, and severe risk) was relatively poor in a study of older hospitalized Australian patients. Prediction modeling using the component domains of the MPI together with additional clinical features and machine learning (ML) algorithms might improve prediction accuracy. Objective: This study aims to assess whether the accuracy of prediction for 12-month mortality using logistic regression with maximum likelihood estimation (LR-MLE) with the 3-category MPI together with age and gender (feature set 1) can be improved with the addition of 10 clinical features (sodium, hemoglobin, albumin, creatinine, urea, urea-to-creatinine ratio, estimated glomerular filtration rate, C-reactive protein, BMI, and anticholinergic risk score; feature set 2) and the replacement of the 3-category MPI in feature sets 1 and 2 with the eight separate MPI domains (feature sets 3 and 4, respectively), and to assess the prediction accuracy of the ML algorithms using the same feature sets. Methods: MPI and clinical features were collected from patients aged 65 years and above who were admitted to either the general medical or acute care of the elderly wards of a South Australian hospital between September 2015 and February 2017. The diagnostic accuracy of LR-MLE was assessed together with nine ML algorithms: decision trees, random forests, extreme gradient boosting (XGBoost), support-vector machines, naïve Bayes, K-nearest neighbors, ridge regression, logistic regression without regularization, and neural networks. A 70:30 training set:test set split of the data and a grid search of hyper-parameters with 10-fold cross-validation-was used during model training. The area under the curve was used as the primary measure of accuracy. Results: A total of 737 patients (female: 370/737, 50.2%; male: 367/737, 49.8%) with a median age of 80 (IQR 72-86) years had complete MPI data recorded on admission and had completed the 12-month follow-up. The area under the receiver operating curve for LR-MLE was 0.632, 0.688, 0.738, and 0.757 for feature sets 1 to 4, respectively. The best overall accuracy for the nine ML algorithms was obtained using the XGBoost algorithm (0.635, 0.706, 0.756, and 0.757 for feature sets 1 to 4, respectively). Conclusions: The use of MPI domains with LR-MLE considerably improved the prediction accuracy compared with that obtained using the traditional 3-category MPI. The XGBoost ML algorithm slightly improved accuracy compared with LR-MLE, and adding clinical data improved accuracy. These results build on previous work on the MPI and suggest that implementing risk scores based on MPI domains and clinical data by using ML prediction models can support clinical decision-making with respect to risk stratification for the follow-up care of older hospitalized patients.