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

At a Glance

Key Highlights

• Seven algorithms were benchmarked; logistic regression was selected for better test-set generalization, the lowest test-set Brier score among the models, and more consistent calibration.
• The ten-variable nomogram includes NIHSS score, Barthel index, LDL cholesterol, sleep quality, and income level.
• Variable reduction proceeded from univariate screening to LASSO, narrowing 49 candidate variables to 18; SHAP-informed interpretation, redundancy, cost, and bedside feasibility informed the final ten-variable selection.
• The test partition contained approximately 36 patients; reported sensitivity and specificity are compatible with 13 of 15 PSA cases and 20 of 21 non-PSA cases.
• Evaluation partitions came from the same multicenter cohort and recruitment window; the reported results are internal validation, not external validation.

Guideline-Based Recommendations

Diagnosis

• The article describes prediction of PSA and does not provide diagnostic criteria for PSA.

Management

• The correspondence describes the nomogram as a candidate instrument, not a bedside decision-support tool ready for use pending external validation.
• The correspondence identifies prospective external studies as the appropriate next step.

Monitoring & Follow-up

• The correspondence cites guidance that calibration is important for interpreting predicted probabilities and that external validation should use data differing from the development sample by geography, time, or case-mix, with sufficient sample size to estimate performance precisely.

Risks

• The small test partition makes performance estimates sensitive to individual classification errors; one additional false negative would reduce sensitivity from 0.867 to 0.800.
• The correspondence reports inconsistencies in cohort descriptions, PSA prevalence, cross-validation folds, Brier scores, and stacking-model AUCs across the text, tables, and figures.
• SHAP attributions describe contributions to model output and do not establish causal effects or selection validity.

Patient & Prescribing Data

The model was developed and evaluated using records from a ten-centre Chinese cohort; the test partition contained approximately 36 patients.

No treatment or prescribing regimen is described. The correspondence characterizes the nomogram as a candidate instrument, not a decision-support tool ready for bedside use pending external validation.

Clinical Best Practices

• The correspondence argues that calibration and clinical deployability can inform model selection when standard performance is comparable.
• Distinguish SHAP-based interpretation from variable-selection methods, as described in the correspondence.
• Distinguish internal evaluation from external validation; the correspondence notes that Figure 1 inaccurately labels the test set as external and independent.
• The correspondence calls for correction of discrepancies in cohort counts, prevalence, validation methods, and performance metrics.

Related Resources & Content