📄 ALBA: An Automated Framework for Benchmarking Clinical Language Biomarkers against Standardized Corpora

Summary

Automated software can now extract many quantitative measures from patients' connected speech. Clinicians still struggle to interpret an individual score without norms produced by the same methods. The authors present ALBA (Automated Language Biomarker Application), a web tool built into the Open Brain AI platform. It compares one patient's language measure with reference distributions drawn from standardized clinical corpora, mainly AphasiaBank and other TalkBank repositories. The reference values and the clinician's data come from the same elicitation tasks (e.g., cookie theft, Cinderella story) and the same Open Brain AI processing pipeline, so the comparison is not distorted by differences between tools. ALBA does not reprocess transcripts. It stores precomputed summary statistics for each task–diagnosis–measure combination and then automatically chooses either a one-sample t-test or a Wilcoxon-based non-parametric test, depending on whether the reference distribution is normal. Output includes a p-value, an APA-style textual interpretation, and box, bar or violin plots. The tool makes normative benchmarking of discourse measures quick and accessible, supporting differential diagnosis, therapy planning and baseline tracking.

Key points

• Reference cohorts include healthy controls and several clinical groups: aphasia subtypes, LHD, RHD, TBI, MCI and dementia. Norms are derived from more than 600 transcripts sourced from AphasiaBank/TalkBank.
• Each Task × Diagnosis × Linguistic Variable cell stores eleven precomputed metrics, including mean, SD, median, min/max, Q1/Q3, Shapiro-Wilk p-value and skewness. Together these set an IsNormal flag.
• Test choice works as follows:
  • A t-test is used if Shapiro-Wilk p > .05, or if n ≥ 30 with low skewness.
  • A Wilcoxon signed-rank approximation is used otherwise, including when |µ − M| > 0.1µ.
  • Results are labeled "Significant" or "Similar" at α = 0.05.
• Computation runs client-side with the jStat library, so the patient's value never leaves the browser.
• Aggregated norms for 290 linguistic biomarkers are released as a machine-readable CSV via openbrainai.com. Raw transcripts remain under TalkBank's registration-based access agreement.