Source-linked AI summary
Predicting Multiple Clinical Outcomes Related to Functional Recovery and Social Isolation Among Older Adults After Lower-Limb Fracture or Hip Replacement
Santosh Ray, Pratik K. Mishra, Ali Abedi, Charlene H. Chu, Amir Ahmad, Shehroz S. Khan
TL;DR
Older adults recovering after lower-limb surgery may face intertwined mobility and social-engagement challenges that are not routinely assessed together. This study formulates simultaneous prediction of five clinical scores from multimodal sensor data as multi-output regression and finds joint deep-learning models outperform separate and traditional approaches.
Problem
Recovery can involve both mobility and social-engagement limitations, while regular assessment of these dimensions is mostly not covered by primary care services.
Method
Using daily multimodal sensor features from 18 participants monitored over 1,008 days, the study compares traditional machine-learning and tabular deep-learning single- and multi-output regressors for five clinical measures.
Results
Multi-output regressors performed better than single-output regressors across tested machine-learning and deep-learning approaches, with NODE emerging as the best multi-output regressor.
Takeaways & Limitations
Feature importance suggests that mobility and physiological data are important for estimating functional recovery and social isolation jointly.
Takeaways & Limitations
The study analyzed multimodal sensor data from only 18 patients, with a female-dominated cohort and a single geographic region that included potentially confounding seasons.
Abstract
from arXiv · showhide
Older adults recovering after lower-limb fracture or hip replacement may experience complex recovery trajectories. Most of the time, these clinical aspects are studied in isolation, masking their joint impact on recovery. This study used the MAISON-LLF dataset, which contains multimodal sensor and clinical assessment data from 18 older adults recovering in the community after lower-limb fracture or hip replacement. Participants were monitored for up to eight weeks, corresponding to a maximum of 1,008 participant-days of sensor monitoring. Forty-six daily features were extracted from indoor motion, acceleration, step count, heart rate, out-of-home mobility, and sleep data. Five clinical outcomes were assessed every two weeks: the Social Isolation Scale, Oxford Hip Score, Oxford Knee Score, Timed Up and Go test, and 30-second Chair Stand test. We utilize an inherent relationship between multi-modal sensor data and different clinical scores and formulate it as a multi-output regression problem. We tested various machine learning and deep learning single- and multi-output regression algorithms to predict these scores simultaneously. The results showed that predicting clinical scores jointly was better than separately. The tabular DL multi-output regressor, NODE, gave a remarkable performance of MSE=3.96 and MAE=1.02 in comparison to other multi- and single-output regressors. The SHAP feature analysis further showed the importance of including multimodal sensors to provide a good estimate of patients' recovery trajectory. This work may support the simultaneous assessment of functional recovery and social engagement among community-dwelling older adults and ultimately help improve their care and quality of life.
I. INTRODUCTION
Recovery after lower-limb fracture or hip replacement can affect mobility and social engagement, but these dimensions are rarely assessed together. This study formulates simultaneous prediction of five clinical measures from multimodal sensor data.
- Community recovery can involve limitations in both mobility and social engagement, while routine primary care often does not regularly assess both dimensions.
- Mobility, sleep, heart rate, and other physiological indicators have been separately associated with functional mobility and social isolation.
- Multimodal data spanning indoor motion, outdoor movement, sleep, heart rate, and steps can provide an overall profile for predicting functional decline and social isolation.
- The MAISON-LLF dataset contains longitudinal wearable, ambient, sleep, mobility, and clinical assessment data from older adults recovering in the community.
- The study formulates prediction of OHS, OKS, TUG, Chair Stand, and SIS as a multi-output regression problem to model these outcomes jointly.
- Tabular deep learning models were investigated, while feature selection and SHAP analysis examined whether multimodal sensor variables were useful for prediction.
II. RELATED WORK
Prior work has used sensor-based machine learning to monitor social, functional, and rehabilitation-related outcomes in older adults. Studies specifically addressing lower-limb fracture recovery after discharge remain limited.
- Sensor-based monitoring has been applied to loneliness, social isolation, social interaction, home visits, frailty, activity, and rehabilitation-related outcomes.
- Prior studies have used wireless, ambient, wearable, smartwatch, inertial, and insole sensors with machine-learning models for older-adult monitoring.
- Studies closer to lower-limb fracture recovery include wearable insole sensors for predicting fracture-healing progression and multimodal MAISON data for post-discharge recovery research.
B. MULTI-OUTPUT AND MULTI-TASK LEARNING IN HEALTH-RELATED PREDICTION
Multi-output and multi-task learning have been used for related health, biomedical, and sensor-prediction problems, but post-discharge recovery after lower-limb fracture remains underexplored. The paper addresses this gap by jointly modelling functional and social recovery outcomes.
- Health-related multi-task studies have jointly predicted clinical outcomes from neuroimaging, biomedical measurements, and longitudinal multimodal data.
- EHR-based multi-task learning has addressed mortality, decompensation, length of stay, phenotype, and other clinical prediction targets.
- Sensor-based multi-task learning has jointly addressed activity, user, behavioral, and composite activity recognition tasks.
- Earlier sensor studies usually defined one target at a time, including loneliness, social isolation, frailty, activity, fracture healing, or recovery status.
- Prior multi-output and multi-task studies largely focused on other domains, while tabular deep-learning models received limited evaluation for this recovery setting.
- The proposed setting suits multi-output learning because SIS, OHS, OKS, TUG, and related measures represent dimensions of the same post-discharge recovery process.
III. METHODS
The study uses longitudinal multimodal sensor and clinical data from community-dwelling older adults recovering after lower-limb fracture or hip replacement. Daily sensor features were linked to repeated clinical assessments over an eight-week monitoring period.
- The MAISON-LLF dataset followed older adults from shortly after discharge for up to eight weeks using longitudinal multimodal sensor and clinical assessment data.
- Clinical assessments included hip and knee pain/function scores, functional mobility and balance, and lower-extremity strength and capacity.
- 1,008 days of multimodal monitoring came from 18 participants with daily sensor-derived features and repeated questionnaire and physical-performance assessments.
- The cohort had a mean age of 76.5 years, included 14 female participants and 14 participants living alone, and represented hip replacement and several fracture types.
B. EXPERIMENTAL SETUP
The study compares traditional machine-learning and tabular deep-learning regressors for multimodal clinical-score prediction, using participant- and week-based validation to assess generalization.
- The evaluation included LR, ET, RF, KNN, SVR, NODE, FT-Transformer, TabPFN, and TabNet regressors.
- Models used min-max scaling to align target ranges before training and inverse transformation before computing MSE and MAE.
- Leave-one-person-out cross-validation tested generalization to completely unseen older adults without prior exposure to their baseline behaviours.
- Leave-one-week-out cross-validation tested robustness to week-to-week behavioural shifts across the cohort.
A. CORRELATIONS AMONG CLINICAL OUTCOMES
Clinical outcomes showed coupled relationships between functional mobility and social isolation, with key correlations remaining stable across unseen participants.
- TUG exhibited strong negative correlations with Chair Stand (-0.49) and SIS (-0.43) in the overall dataset.
- The overall metrics indicate that functional decline and social isolation may not occur independently.
- The negative TUG–Chair Stand correlation persisted across participant folds, ranging from -0.43 in Fold 9 to -0.55 in Fold 10.
- Stable correlations across 18 participant folds support jointly predicting the five clinical outcomes rather than isolating each target.
B. LEAVE-ONE-PERSON-OUT CROSS-VALIDATION
Leave-one-person-out validation showed that multi-output prediction generally outperformed single-output prediction, with the strongest gains from tabular deep-learning models, especially NODE.
- p-value = 0.006 for the multi-output performance improvement over single-output regression.The comparison used Wilcoxon signed-rank testing across leave-one-person-out results.
- NODE reduced average MSE from 36.16 in single-output regression to 19.07 in multi-output regression.
- Traditional models showed smaller or negative gains, including RF’s 8.49% average MSE improvement and a -5.2% SIS performance change.
- 50.54% was NODE’s overall average MSE improvement and 45.56% its average MAE improvement when switching to multi-output prediction.
- 88.77% was NODE’s average MSE improvement across all five variables relative to the multi-output LR baseline.
- 78.04% was NODE’s average MAE improvement relative to the multi-output LR baseline, compared with 19.32% for RF and 16.65% for ET.
C. LEAVE-ONE-WEEK-OUT CROSS-VALIDATION
Leave-one-week-out validation tested whether models generalized across unseen temporal behavior. Multi-output tabular deep learning models generally benefited most from jointly predicting the five clinical outcomes.
- 6.05 MSE was achieved by multi-output FT-Transformer, compared with 26.01 for its single-output counterpart.
- 2.38 MSE was achieved by multi-output NODE, improving from 4.8 in the single-output setting.
- 0.50 average MAE was achieved by multi-output NODE, down from 0.88 with single-output prediction.
- 63.51% overall MSE improvement and 55.74% MAE improvement made FT-Transformer the strongest relative multi-output gain.
- 96.36% average MSE improvement over the multi-output LR baseline was achieved by NODE across the five clinical variables.
- 90.66% average MAE improvement over the multi-output LR baseline was achieved by NODE, followed by FT-Transformer at 80.78%.
D. FEATURES DISTRIBUTION ANALYSIS
Sensor features showed strongly skewed distributions and many upper outliers across activity, mobility, heart-rate, and sleep modalities. These patterns reflect predominantly home-based behavior alongside episodic activity, exertion, or disrupted rest.
- Acceleration and motion features were right-skewed, with dense upper outliers and long tails toward 1.0.The pattern is described as prolonged sedentary behavior punctuated by bursts of activity or nighttime restlessness.
- Step-count, step-mean, and step-max features displayed massive densities of upper outliers.
- Position-duration, position-distance-travelled, and position-count had medians near the distribution floor, with most variation appearing as extreme upper outliers.The cohort spent the vast majority of time at home.
- Heart-rate maxima showed upper outliers, while several sleep features had extreme upper-bound values representing episodic exertion or disrupted rest.
E. FEATURES IMPORTANCE AND INTERPRETABILITY
SHAP analysis of NODE identified behavioral timing and sleep characteristics as more informative than raw activity volume. Several volumetric and outdoor-mobility measures ranked near the bottom, partly because outside movement was sparse.
- motion-max-timestamp was the most important feature, with an overall average rank of 4.11.The timing of maximum daily motion was more predictive than raw physical-activity volume.
- sleep-deep ranked third overall at 7.61, while sleep-rem and sleep-duration-to-wakeup ranked within the top ten.
- Step-count ranked 37.33, acceleration-sum 38.83, and acceleration-movement-events-24h 40.72, placing raw activity counters near the bottom.
- Position-distance-travelled ranked last at 43.89, because outside positional variation was too sparse to generalize reliably across folds.
V. CONCLUSION AND FUTURE WORK
The study found that multi-output regression outperformed single-output approaches, with deep learning and NODE performing best, while multimodal mobility and physiological features supported prediction of recovery and social isolation. Interpretation is constrained by the small, geographically limited cohort and repeated daily records derived from only 18 participants.
- Multi-output regressors outperformed single-output regressors across the tested machine learning and deep learning approaches.
- The dataset comprised 1,008 monitoring days from only 18 patients, limiting the independent sample size.
- Future work includes longitudinal individual recovery profiles, sequential deep learning models, richer phenotype descriptions, and geographic or neighborhood context.