A Random Forest model using pretreatment clinical data and initial rehabilitation/aquatic-therapy plans predicted insufficient motor improvement (ΔFMA <10) after aquatic therapy in stroke with temporal-validation AUC 0.818, and adding actual day-14 treatment-execution data did not improve prediction; external validation and prospective clinical-impact evaluation are required before clinical use.
About one-third of stroke patients in this single-center cohort had insufficient motor improvement after aquatic therapy.
Random Forest was the best-performing model for Model A (baseline plus initial treatment plan) in temporal validation, with good discrimination.
At a frozen sensitivity-prioritized threshold, Model A had high sensitivity and negative predictive value but modest specificity and PPV.
Adding actual treatment-execution parameters accumulated through day 14 (Model B) did not improve prediction over Model A.
SHAP analysis of Model A identified planned aquatic-therapy parameters, baseline motor function and time since stroke as the most important predictors.
Model B illustrates how risk could be updated at day 14 once real execution records are available, but the study does not show that model-guided adjustment improves recovery.
The study used temporal validation and multiple evaluation approaches beyond discrimination.
This study built computer models to predict which stroke patients are unlikely to gain meaningful motor recovery (a gain of fewer than 10 points on the Fugl-Meyer scale) after about eight weeks of aquatic therapy. Using records from 1,320 patients at one hospital, about 31% had insufficient improvement. A Random Forest model that used information available before the first pool session, such as baseline motor scores, time since stroke, and the planned length and frequency of aquatic therapy, identified these patients fairly well when tested on a later group of patients (AUC 0.818). At the chosen cutoff, the model caught most patients who would not improve enough (sensitivity about 86%) and was usually right when it said someone was low-risk (about 91%). However, about half of those flagged as high risk actually improved sufficiently. Adding data on the therapy actually delivered during the first two weeks did not make predictions better, though that version could in principle be used to update risk at day 14. This research suggests that routinely collected information could help flag patients who might need a different rehabilitation approach. However, the study was retrospective and from a single center, and it did not test whether using the model to change treatment improves recovery. The authors say external validation and prospective studies are needed before any clinical use.
Chen Y, Zhang J, Qie S. (2026). Development of an interpretable machine-learning model to predict insufficient functional improvement after aquatic therapy in stroke.. Frontiers in neurology. https://doi.org/10.3389/fneur.2026.1871732