Explainable Prediction of Poststroke Anxiety: Translating Model Accuracy into Clinical Practice

By
Jorge Valle-Reconco
Patricia Soriano
Isaac Zablah
October 15, 2026

Digital Health

Overview

A multicenter study evaluated seven machine-learning algorithms for predicting poststroke anxiety (PSA); logistic regression was selected based on test-set performance, Brier score, and calibration, and was presented as a ten-variable nomogram. The reported test partition was small and drawn from the same pooled cohort as the training data, so the model has internal rather than external validation and is not ready for bedside decision support.

Background

Poststroke anxiety is a clinically relevant emotional disorder that can affect functional rehabilitation, quality of life, and longer-term outcomes. Its contributing factors include neurobiological changes, cognitive impairment, behavioral responses, and social circumstances. The source describes a prediction model intended to combine neurological, metabolic, and psychosocial information in a bedside-oriented format. It also emphasizes that discrimination alone is insufficient to establish clinical usefulness: calibration and validation in populations distinct from the development cohort are important considerations.

Data Highlights

Measure or feature	Reported result
Algorithms benchmarked	7
Training-set AUC	Several algorithms achieved 1.000
Selected model	Logistic regression; test-set AUC 0.981 and lowest test-set Brier score among the seven models
Candidate variables	49 screened; reduced to 18 by univariate screening and LASSO regression
Nomogram variables	10, selected with consideration of SHAP importance, redundancy, cost, and bedside feasibility
Test partition	Approximately 36 patients: 15 PSA and 21 non-PSA, as inferred from reported sensitivity and specificity
Test-set classification	Sensitivity 0.867 (13/15); specificity 0.952 (20/21)
Validation status	Internal validation; partitions came from the same multicenter cohort and recruitment window

Key Findings

Seven algorithms were compared; although several had training-set AUCs of 1.000, logistic regression was chosen because it generalized better in the held-out test set, had the lowest test-set Brier score, and showed more consistent calibration.
Variable reduction proceeded from univariate screening to LASSO, narrowing 49 candidate variables to 18; SHAP was then applied to the fitted logistic regression model, with its importance ordering considered alongside redundancy, cost, and bedside feasibility when selecting ten nomogram variables.
The nomogram includes NIHSS score, Barthel index, LDL cholesterol, sleep quality, and income level, among its ten variables.
SHAP attributions describe contributions to the fitted model output and do not establish variable-selection validity or causal effects; the source characterizes the final variable choices as interpretation-informed simplification, not selection by SHAP alone.
The test partition contained approximately 36 patients. The reported sensitivity corresponds to 13 of 15 PSA cases and specificity to 20 of 21 non-PSA cases; one additional false negative would reduce sensitivity to 0.800.
The source identifies inconsistencies in the published record: Figure 1 labels the test partition as an external independent cohort despite pooled random allocation from ten centres, and reported PSA/non-PSA counts and prevalence differ across Figure 1, the text, and Table 3.

Clinical Implications

The source describes the nomogram as a clinically motivated candidate instrument, not a decision-support tool ready for bedside use, pending prospective external validation. Separately, the supplied guidance context states that anxiety screening after stroke should use validated tools and that screening may be revisited at care transitions or changes in health status; the Canadian recommendations also state that psychological interventions such as cognitive-behavioral therapy should be considered.

Conclusion

The study presents an explainable, bedside-oriented PSA prediction model and gives calibration and usability a role in model selection. Its small internal test partition and reporting inconsistencies, together with the absence of external validation, limit conclusions about transportability and bedside readiness.