Crownlands - Clinical Assessment and Outcomes Voice Data (Preview)

Audio from clinician-patient visits with structured interviews → clinician’s assessment of the patient.

What to know: This dataset is a trove of long-duration patient visits where a doctor assesses a patient’s health in a structured way. Along with real-world medical conversation, these interactions have hard outcomes. In patient care and in clinical trials, clinical scoring systems are used to standardize diagnostic staging and measure longitudinal progression. This dataset contains recorded visits where a physician scores patients on these metrics. This is a task that should be trained - correct measurement requires the application of clinical judgement, and the administration and grading of these cognitive assessments are near-term steps towards AI doctors holistically diagnosing patients.

Training Goal: This dataset trains (voice) models on clinical conduct with patients and how physicians assess patient symptoms and behavior. Improvement on the clinical assessment task requires careful objective scoring of responses and holistic clinical judgement.

Contents:

With each available clinical visit:

Three speakers - doctor, patient, partner (informant)
1-2 hours duration per visit
Three independent assessments - see below

Each recording contains three clinical/cognitive assessments, conducted as structured interviews. Two of these tasks are scoring tasks, which have multiple yes/no/choice questions. The longest task is a clinical judgement task, which follows a structured interview but relies un the clinician’s expert judgement to determine the final symptom rating.

Task medical details:

Scoring Task A: NPI-Q This is the hidden text or content that drops down.


Scoring Task B: GDS This is the hidden text or content that drops down.


Clinical Judgement Task: CDR This is the hidden text or content that drops down.


Health Data Provenance and Privacy:

This dataset contains de-identified data from real patients discussing their own cases with a physician. All data is shared safely and with patient consent. Crownlands cares about transparency of the data chain-of-custody.

In this dataset, the human health data is “first-party” - meaning Crownlands sponsored and generated the data and is providing the data to users.

Per Crownlands policy, all users must agree not to attempt re-identification of the patients or speakers in order to access the dataset.

HUMAN DATA PROVENANCE

Patient consent  .----. ___ .----.   Crownlands serves
in Crownlands   (     ()   ()     )  de-ID data
study            `----' ‾‾‾ `----'