Item new Preprint · not peer-reviewed Added 31 October 2026

Respiratory infection hospitalisation risk in New Zealand predicted by gradient boosting

Gradient boosting models using demographics, morbidity indices and influenza vaccination status separated people with and without a respiratory infection hospitalisation, across about 5 million New Zealanders each year.

The study modelled annual respiratory infection hospitalisation in 4.85 to 5.36 million New Zealanders, 2017 to 2024. Six algorithms used demographics, hospital and pharmaceutical morbidity indices and vaccination status, with cross-validation grouped by demographic profile. Annual prevalence was 1.01% to 1.60%; LightGBM mean AUROC was 0.8430 and XGBoost 0.8415. Morbidity indices raised LightGBM average precision over demographics from 0.0720 to 0.1158, and vaccination status to 0.1283. Trained on one year and tested on the next, LightGBM averaged AUROC 0.8429 and XGBoost 0.8423. SHAP linked vaccination to lower estimated risk mostly in the oldest group, which describes model behaviour, not vaccine effect.

Why it is interesting: Routine hospital and pharmacy records added most of the gain over demographics alone. Performance held when models trained on one year were tested on the next.

Source: medRxiv, 26 October 2026
Paper: Population-scale machine learning models to estimate individual-level risk for acute respiratory infection
Authors: Albrecht S, Smith E, Yeo HY, Park JY, Gul A, Hills I et al.
DOI: 10.64898/2026.10.23.26366351
Type: Preprint
Design: Retrospective national linked administrative data, ~4.85-5.36 million people per year, 2017-2024; six ML algorithms, eight variable sets, five-fold grouped cross-validation plus next-year validation
Verdict: New finding
Driver: What linked data reveals
Driver: AI in medicine
Driver: Infection and pandemics

Cite this item

Other drivers in Data, models and systems