Answers are grounded in this paper and cite their sources.
Complex hospital environments and digital healthcare systems may create healthcare access challenges for older adults, highlighting the potential need for medical visit companion services (MVCS). However, empirical research on disaggregated MVCS needs and their associated features remains scarce. This study aimed to apply interpretable machine learning algorithms to identify factors associated with MVCS needs and elucidate the non-linear effects of key features. A cross-sectional study was conducted among 494 older adults in Zhejiang Province, China, from January to May 2024. MVCS needs were quantified using a validated 15-item scale with established reliability and validity. Data were randomly divided into a training set (70%, n = 347) for 10-fold cross-validation and a test set (30%, n = 147). Seven machine learning regression algorithms, including linear regression, generalized additive model (GAM), random forest regression, extreme gradient boosting regression (XGBoost), support vector regression, multilayer perceptron regression, and k-nearest neighbors regression, were trained and benchmarked against a naive mean-only baseline. For the selected algorithm, feature importance was quantified in the training set using gain-based metrics and mean absolute SHAP values. Accumulated Local Effects (ALE) plots were generated using the training set to characterize the non-linear feature-response profiles of the key features. The mean total MVCS needs score among older adults was 56.36 ± 12.13, with the service quality assurance dimension scoring the highest. Notably, 67.8% of older adults reported limited familiarity with MVCS. Although XGBoost yielded the most favorable numerical performance metrics (test-set R 2 = 0.439, 10-fold cross-validation R 2 = 0.412 ± 0.108) across algorithms, bootstrap comparisons indicated no statistically significant performance advantage relative to the other algorithms. Both XGBoost gain and mean absolute SHAP values consistently ranked family care, social support, individualized companionship demand, and MVCS awareness among the most important associated features of MVCS needs. However, sensitivity analyses could not establish the indispensability of individualized companionship demand. ALE values for family care and social support were positive relative to the mean-centered reference at low-to-moderate feature values, changing only modestly there before declining into negative values at higher levels. By contrast, ALE values for individualized companionship demand and MVCS awareness generally shifted from negative at low scores to positive at higher scores. Older adults reported substantial needs for MVCS, with service quality assurance ranking the highest among all assessed dimensions. Nevertheless, given the current limited awareness of MVCS, these needs may reflect expectation-based potential demand rather than realized needs. Although it yielded no statistically significant performance advantage over alternative algorithms, XGBoost was adopted for subsequent feature importance analyses given its built-in capacity for internal gain-based evaluation and robust integration with post-hoc interpretability frameworks. Family care, social support, and MVCS awareness emerged as consistently leading features, while individualized companionship demand also ranked highly in the primary analysis but requires more cautious interpretation. These features exhibited nonlinear associations with MVCS needs, although patterns in sparsely populated feature intervals should be interpreted cautiously. This algorithm serves solely for exploratory association analysis and interpretation; it is not designed for real-world screening or operational use. Not applicable.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: