Assessing Clinical Readiness for Sepsis Prediction Using Vital Signs

By
Li Li
Song Lei
Chansokhon Ngan
October 27, 2026

Jmir Medical Informatics

Objective:

Evaluate methodological and interpretational issues in Yang et al.’s vital-sign-based model for online sepsis prediction.

Approach:

Data selection: The commentary examines the reduction from 40,336 PhysioNet records to 3,766 sequences after vital-sign completeness filtering and asks how residual missing values were handled.
Model evaluation: It questions how margin and epoch selection related to the test data, and how the reported 4:1 train/test split related to 5-fold cross-validation.
Prediction timing: It reviews the reported normalized MAE and sensitivity and specificity, noting that the timing metric and classification threshold need clarification.

Key Findings:

Completeness filtering may favor records with more complete temporal data; the commentary notes this may not reflect settings with substantial data sparsity.
The manuscript does not clearly describe residual-missingness preprocessing or the relationship between model selection, test-set assessment, and cross-validation.
The abstract reports normalized MAE as 0.11%, whereas Table 2 lists 0.11; if 0.11 is a proportion, it corresponds to 11% of sequence length.
The reported normalized MAE does not directly indicate prediction timing in hours, and the probability threshold for sensitivity of 89.29% and specificity of 73% is not stated.

Interpretation:

The commentary requests clearer reporting of missing-data handling, validation procedures, prediction timing, and the classification threshold to support assessment of reproducibility, transportability, and clinical readiness.