Background: Sepsis is a leading cause of pediatric mortality worldwide, disproportionately affecting children in low- and middle-income countries (LMICs). However, timely recognition of potential sepsis and access to healthcare resources needed to diagnose pediatric sepsis according to international guidelines are challenging in LMICs. This exploratory study aimed to develop machine learning (ML) models to detect pediatric sepsis and septic shock using a simplified set of clinical data contextualized for practical use in resource-limited settings. Methods: This was a secondary analysis of an observational study of 100 children with potential sepsis admitted to a non-profit referral hospital in Dhaka, Bangladesh. The outcomes were sepsis as defined by a Phoenix Sepsis Score (PSS) ≥ 2 and septic shock (sepsis plus PSS cardiovascular sub-score ≥ 1). Models were trained using either clinical + laboratory variables or clinical-only variables. A single 24 h worst-value assessment window was derived per patient; stratified 5-fold cross-validation was used to maintain class proportions across the training and test folds. Model performance was assessed using area under the precision–recall curve (AUPRC) and area under the receiver operating characteristic curve (AUROC) with 95% confidence intervals (CIs) derived from a 2000-resample patient-level bootstrap of out-of-fold classifications. Logistic regression coefficients were used to assess feature contributions. Results: For sepsis classification, the non-laboratory model achieved an AUPRC of 0.942 (95% CI: 0.884–0.979) and an AUROC of 0.945 (95% CI: 0.890–0.983), with comparable performance from the clinical + laboratory model (AUPRC 0.941, 95% CI: 0.880–0.981; AUROC 0.945, 95% CI: 0.881–0.986). For septic shock, AUROCs of 0.870 (95% CI: 0.761–0.952) and 0.878 (95% CI: 0.758–0.967) were observed. However, these estimates should be interpreted cautiously, given the low prevalence (23%) and absence of external validation. SpO2:FiO2 ratio, GCS, and systolic blood pressure were consistently strong predictors across models. Conclusions: ML models using pragmatic clinical variables demonstrate preliminary diagnostic performance, with the non-laboratory model showing discrimination comparable to models incorporating laboratory data. Logistic regression demonstrated the most stable performance and may represent an early proof of concept for assistive diagnostic support. However, these models are not clinically usable without external validation. These findings are hypothesis-generating; external validation in larger, independent cohorts is essential before any clinical use, particularly for septic shock.
BACKGROUND Accurate emergency department (ED) volume forecasts support staffing, capacity planning, and budgeting, but conventional time-series forecasting can require specialized analytic expertise. Consumer-facing large language model (LLM) interfaces may lower this barrier, although their forecasting accuracy, consistency, and sensitivity to data recency are uncertain. OBJECTIVE To compare forecasts produced by commonly available generative AI interfaces with a Holt-Winters exponential smoothing baseline and to determine whether progressively updating the historical data window (“proximal forecasting”) identifies a pragmatic recency point for fiscal-year planning. METHODS We conducted a two-part retrospective observational forecasting study using monthly ED visit volumes from four hospitals. In part 1, 12-month forecasts were generated from 12-, 24-, and 36-month historical windows using Holt-Winters and five LLM interfaces (ChatGPT, Claude, Copilot, Perplexity, and Gemini); LLM forecasts were repeated 10 times to characterize run-to-run variability. In part 2, ChatGPT and Gemini were evaluated using 24-month windows whose end dates were systematically reduced from 7 months before to immediately before FY2025 and were compared with an R-implemented Holt-Winters baseline. Accuracy was assessed using mean absolute percentage error (MAPE) and root mean square error (RMSE). RESULTS Across all four sites, Holt-Winters had the lowest aggregate MAPE with a 24-month historical window (3.26±0.68) compared with 12 months (4.29±0.46) and 36 months (3.43±0.73). With 24 months of history, two LLM interfaces (ChatGPT 5.5 and Gemini 3.1) produced MAPE values similar to Holt-Winters at ED A (2.17±1.22 and 2.10±1.04 vs 2.56±1.08) and ED B (2.35±1.16 and 2.21±1.21 vs 2.46±1.39), but had higher error at the pediatric ED (7.35±5.26 and 7.29±5.27 vs 5.90±2.05) and the summer-seasonal community ED (8.74±3.11 and 8.69±3.11 vs 2.12±1.35). No approach met a predefined threshold in the Ped ED and community ED. In adult EDs with typical seasonality, forecasts generated up to three months from the fiscal-year start were comparable to fiscal-year–proximal forecasts (mean MAPE difference -0.4%–1.8%). Pediatric and highly seasonal sites demonstrated higher baseline error and recency bias susceptibility. CONCLUSIONS Selected LLM interfaces can produce monthly ED volume forecasts comparable to Holt-Winters in some adult ED settings when given sufficient historical context. A forecast update up to 3 months before the fiscal year may be a pragmatic planning point for sites with typical seasonality. Pediatric and strongly seasonal EDs may require local validation and tailored models.
30 August 2026 in African Journal of Emergency Medicine