AI-based prediction of SPPB scores using questionnaires of abilities: findings from the national health and aging trends study

Abstract

BACKGROUND: The Short Physical Performance Battery (SPPB) is a widely used assessment tool to evaluate lower extremity function in older adults. However, it requires clinical settings which may not be feasible in all circumstances. This study aimed to develop alternative methods for indirectly estimating SPPB scores using questionnaire responses related to functional abilities. METHODS: We analyzed data from Round 12 of the National Health and Aging Trends Study, using 4,988 participants for statistical analyses, and 2,035 participants (1,628 for training and 407 for testing) for model development and validation. A total of 27 questionnaire items, covering basic and instrumental activities of daily living and physical activities, were used as predictors. Three artificial intelligence models were developed: a tree-based classifier, a multilayer perceptron (MLP) classifier, and a tree-based regressor. For comparison, summed abilities of each ability category and simplified summed ability derived from Shapley Additive Explanations analysis were used. Multiclass and binary classifications were performed using predefined SPPB cutoff values (scores ≤ 3 and ≥ 10). RESULTS: In analysis comparing SPPB score groups (0-3, 4-9, 10-12), all 27 questionnaire variables were statistically significant. The summed abilities showed a Pearson correlation of 0.716 with total SPPB scores. In multiclass classification, the MLP classifier outperformed other models with a mean AUC of 0.803 (95% CI: 0.767-0.839). For binary classification, distinguishing between individuals with severe impairment (SPPB ≤ 3) and unimpaired function (SPPB ≥ 10), the MLP classifier again demonstrated the highest AUCs (0.907 for SPPB ≤ 3; 0.920 for SPPB ≥ 10). Summed abilities outperformed AI models in detecting severe impairment, with the total ability score reaching an AUC of 0.915. However, for detecting unimpaired function, AI models consistently outperformed summed abilities (maximum AUC of 0.898). CONCLUSIONS: The proposed AI methods enable prediction of SPPB component scores, supporting indirect functional assessment when SPPB testing is not feasible. These tools can help reduce unnecessary clinical burden and cost by guiding SPPB administration decisions.

Ähnliche Arbeiten