Design: Systematic review and meta-analysis of prediction model studies
Population: 13 studies in Chinese patients with diabetes
Primary outcome: Discrimination (AUC) for hypoglycaemia
Effect: Pooled AUC 0.90 (95% CI 0.87-0.93); logistic regression 0.83

This meta-analysis pooled 13 studies of machine learning models built to predict hypoglycaemia in Chinese patients with diabetes. The pooled prevalence of hypoglycaemia in the source populations was 25%.

The pooled area under the curve was 0.90 (95% CI 0.87 to 0.93). Gradient boosting and random forest models scored highest, but ordinary logistic regression reached 0.83. The predictors the models leaned on were familiar: age, insulin use, body mass index, HbA1c, creatinine and a previous episode.

The authors found frequent methodological weaknesses and little external validation. High discrimination in the population a model was built on often shrinks elsewhere. For now, the useful message is that the models' common predictors are ones a physician already knows.

Ask every patient on insulin or a sulfonylurea about hypoglycaemia since the last visit; a previous episode is a leading predictor.
Review kidney function when judging hypoglycaemia risk, since falling clearance prolongs the action of insulin and sulfonylureas.
Consider relaxing HbA1c targets in older patients on insulin who have had an episode.
Do not rely on a commercial risk score that has not been validated in a population like yours.

Why it matters

The models' common predictors are facts already in the notes, so a careful history remains essential while the models await validation.

The statistics, in plain English

An area under the curve of 0.90 means the model ranks a patient who will have hypoglycaemia above one who will not about nine times in ten. That figure usually falls when a model is tested in a new hospital or population, and most of these models were not.