ML-based Short Physical Performance Battery future score prediction based on questionnaire data

TL;DR

This study explores machine learning models, especially XGBoost, to predict older adults' future physical performance scores from questionnaire data, aiming for early intervention in physical decline.

Contribution

It demonstrates the effectiveness of ML algorithms, particularly XGBoost, in predicting SPPB scores with high accuracy using questionnaire data.

Findings

XGBoost achieved the lowest MAE of 0.79 points.

Feature subset selection maintained prediction accuracy with MAE of 0.82.

ML models can reliably forecast physical performance deterioration.

Abstract

Effective slowing down of older adults\' physical capacity deterioration requires intervention as soon as the first symptoms surface. In this paper, we analyze the possibility of predicting the Short Physical Performance Battery (SPPB) score at a four-year horizon based on questionnaire data. The ML algorithms tested included Random Forest, XGBoost, Linear Regression, dense and TabNet neural networks. The best results were achieved for the XGBoost (mean absolute error of 0.79 points). Based on the Shapley values analysis, we selected smaller subsets of features (from 10 to 20) and retrained the XGBoost regressor, achieving a mean absolute error of 0.82.