Diagnostic accuracy of hospitalization diagnosis codes and exploratory algorithms to identify traumatic brain injury in older adults: a two-centre study

Background — Validated approaches for identifying traumatic brain injury (TBI) in older adults using Ontario health administrative data are lacking.

Methods — We conducted a retrospective validation study using linked electronic patient records (EPRs) from two Level I trauma centres and Ontario administrative data. Adults aged ≥65 years with potential TBI-related encounters (2017–2020) were identified using diagnostic or mechanism-of-injury codes. We independently adjudicated 657 randomly selected EPRs in duplicate and assessed ICD-10 diagnostic performance. Machine learning (ML) models incorporating ICD-10 and physician billing codes informed algorithm development.

Results — Any TBI-related ICD-10 code had 95% sensitivity (95% confidence interval [CI] 90 to 97%) but 76% specificity (95% CI 72 to 80%), and 61% positive predictive value (PPV) (95% CI 55 to 66%). ML models had F1-scores between 0.77 and 0.84. Algorithm 1 (high-specificity ICD-10 codes) achieved 71% sensitivity (95% CI 64 to 77%) and 90% PPV (95% CI 84 to 94%), whereas Algorithm 8 (ML-inclusive) achieved99% sensitivity (95% CI 96 to 100%) but a 42% PPV (95% CI 37 to 47%).

Conclusions — Algorithm 1 may be appropriate for studies prioritizing diagnostic precision, whereas higher-sensitivity algorithms may be preferred for surveillance and applications requiring more complete case ascertainment. External validation is needed to inform broader application.