Development and temporal validation of an interpretable prediction model for delayed diagnosis of benign paroxysmal positional vertigo: a retrospective study from Beijing, China.

OBJECTIVE

To develop and validate an interpretable prediction model for delayed diagnosis of benign paroxysmal positional vertigo (BPPV).

METHODS

This retrospective study included 10,422 BPPV patients between July 2020 and June 2024 in Beijing, China, of whom 2949 (28.30%) had delayed diagnosis (symptom onset-to-diagnosis interval >14 days). Patients were temporally divided into development (n = 7536; training/test split 7:3) and temporal validation (n = 2886) cohorts. Logistic regression (LR), random forest (RF), extreme gradient boosting (XGB), and light gradient boosting machine (LGBM) models were developed and compared.

RESULTS

In the temporal validation cohort, LR achieved the highest discrimination (AUC 0.84, 95% CI 0.82, 0.86), and demonstrated stable performance across datasets. Atypical vestibular symptoms emerged as the strongest predictor of delayed diagnosis. Atypical triggers, motion sickness history, headache history, trauma history, and increasing age were also associated with delayed diagnosis. Firth regression yielded similar estimates. Subgroup validation in participants aged ≥65 years also demonstrated stable performance (AUC 0.83, 95% CI 0.79, 0.86).

CONCLUSIONS

An interpretable prediction model based on routinely collected clinical data was developed to estimate the risk of delayed diagnosis in BPPV. The final LR model demonstrated stable predictive performance across validation cohorts.