Audiograms from 16,000 Vanderbilt patients yield better hearing-loss risk scores than billing codes

"Diagnostic codes are very common because that's the data that's available in large genetic biobanks that facilitate this research," said Andie DeFreese, a clinical audiologist at Vanderbilt and co-first author of the study ^[8][12]. "Using them leads to a kind of gray area in which we're not able to accurately define who has hearing loss and who doesn't." ^[9]

The gray area comes from how genetic biobank studies are usually run. ICD codes sort people into cases and controls, a method the release calls expedient for processing millions of patient charts ^[10]. An audiogram swaps that yes-or-no label for a measured threshold at each tested frequency ^[5]. The team calls the substitution precision phenotyping ^[2]. Scoring the results on an independent validation cohort rules out the simplest way to inflate a polygenic score, which is to grade it on the same people it was fitted to ^[1].

According to the release, the audiogram-based scores showed "significantly superior predictive power" ^[11]. A modest gain and a large one would justify very different investments by a biobank. The summary does not report the accuracy metric, the size of the gap, or how many people were in the validation cohort.

The design also has a tension over the denominator. Audiograms were available for 16,000 people ^[4], while codes exist across millions of charts ^[10]. In a genome-wide study, both cleaner labels and more people add statistical power. I'd want to know whether the code-based score in the comparison was built on the same 16,000 people or on a larger code-defined sample. The first version asks whether a better label helps at equal size. The second asks whether depth beats breadth, the choice a biobank planner actually faces.

"Before symptoms" is in the release's headline, and the team's stated aim is to identify at-risk patients before onset ^[13]. The summary does not describe any test in people whose hearing is still normal. Ranking people better in a validation cohort is a research result. A screening programme also needs scores calibrated to the general population and a treatment to offer the people it flags.

For now the advantage sits with biobanks built the way BioVU is, with de-identified genetic data linked to health records that include full audiograms ^[5].