Can a recording help a model tell apart people who report asthma and people selected as healthy controls? The new Colive Voice analysis gives a model several kinds of information at once. To see what the voice contributes, compare the full model with versions that use just the audio or just the structured information.
- The journal abstract reports an AUROC of 0.83 and a Brier score of 0.18 in a matched sample of 1,218 people.
- In the author manuscript, the two voice tasks together scored 0.71 AUROC. Structured health and demographic data alone scored 0.75.
- The comparison uses self-reported asthma and selected healthy controls. A clinic evaluation would need its own recruitment and reference diagnosis.
- Paper date
- 22 November 2026
- Matched sample
- 1,218 Colive Voice participants
- Journal abstract AUROC
- 0.83
- Journal abstract Brier score
- 0.18
Original publication: 22 Nov 2026 · The date above refers to this brief.
How the model inputs changed ranking performance
AUROC
What the next validation would check
Define the label
Record whether the outcome is self-reported or clinically assessed.
Compare the inputs
Check how performance differs with audio, structured information and both.
Recruit a separate test group
Check ranking and probability errors where the tool would be evaluated.
Look for the comparison that isolates the input
The author manuscript describes two recordings, a sustained vowel and a reading task, alongside structured health and demographic information. In Table 2, the audio-only combination had an AUROC of 0.71, the structured-information model scored 0.75 and the full model scored 0.83. The full score is 0.12 higher than the audio-only score and 0.08 higher than the structured-information score. Those are differences between reported point estimates in the same table.
Source 1 ↗Source 2 ↗Ranking and probability error can favour different versions
AUROC asks whether the model tends to rank a person labelled as a case above a person labelled as a control across possible cutoffs. The Brier score averages squared differences between predicted probabilities and the observed binary labels. Lower scores indicate smaller errors by that measure. In the author manuscript, the full model had higher AUROC than the XGBoost comparator using all three inputs, 0.83 versus 0.81. The comparator had the lower Brier score, 0.17 versus 0.18. The two differences are descriptive and depend on the evaluation data.
Source 1 ↗Source 2 ↗The reference labels and the sample determine the question
The author manuscript lists 609 self-reported cases and 609 matched controls. Each group makes up half of the analysis by design. That can help compare scores across groups. It also means that the observed case proportion and the measured probability error belong to this sample. In a group of patients with similar respiratory symptoms and different diagnoses, a model would face a different set of examples.
Source 1 ↗Source 2 ↗A clinical validation would change the comparison again
An independent study could recruit people for whom the diagnosis is still uncertain and compare predictions with a prespecified clinical assessment. It could report how many people cross a fixed cutoff, how probabilities agree with observed outcomes and how results vary by language or recording conditions. The new report evaluates a candidate research measure. Its matched-sample results give that next study a comparison to work from.
Source 1 ↗Source 2 ↗Read the inputs, labels and scores together
Publisher’s abstract and publication record for the accepted article. Checked the sample size, self-reported reference label and headline performance. The author manuscript provides the detailed comparison.
01What did the journal report for the full model?
- What was observed
- AUROC 0.83 and Brier score 0.18.
- Where the conclusion stops
- These metrics describe performance against the labels in the matched sample.
Source 1 · Published abstract
02Was the highest-ranking model also best by probability error?
- What was observed
- The full model had higher AUROC than XGBoost, while XGBoost had the lower Brier score.
- Where the conclusion stops
- The distinction is between the reported point estimates in this analysis.
Source 2 · Author manuscript, Table 2
Numbers you can inspect
| Measure | Value & unit | Origin & method |
|---|---|---|
| Matched evaluation sample | 1218 people | Reported Study cases and controls together. Source 1 · Published abstract |
| Full model ranking | 0.83 AUROC | Reported Combined voice, health and demographic inputs. Source 1 · Published abstract |
| Full model probability error | 0.18 Brier score | Reported Reported mean squared probability error. Source 1 · Published abstract |
| Self-reported cases | 609 people | Reported Matched case group in the author manuscript. Source 2 · Author manuscript, Table 1 |
| Matched controls | 609 people | Reported Control group in the author manuscript. Source 2 · Author manuscript, Table 1 |
| Two voice tasks only | 0.71 AUROC | Reported Sustained vowel and reading input. Source 2 · Author manuscript, Table 2 |
| Structured information only | 0.75 AUROC | Reported Health and demographic inputs. Source 2 · Author manuscript, Table 2 |
| Full model minus audio-only estimate | 0.12 AUROC units | Calculated Subtract the audio-only score of 0.71 from 0.83. Source 2 · Author manuscript, Table 2 |
| Full model minus structured-only estimate | 0.08 AUROC units | Calculated Subtract the structured-only score of 0.75 from 0.83. Source 2 · Author manuscript, Table 2 |
| XGBoost ranking using all inputs | 0.81 AUROC | Reported Different algorithm using all three input sets. Source 2 · Author manuscript, Table 2 |
| XGBoost probability error using all inputs | 0.17 Brier score | Reported Same comparator and its probability error. Source 2 · Author manuscript, Table 2 |
Compare the actual experiments
These studies answer different questions. Read the unit and endpoint before comparing results.
| Study | Unit & setting | Readout | Interpretation boundary |
|---|---|---|---|
| Combined inputs and a voice-only model Source 2 · Author manuscript, Table 2 | Same matched sample | AUROC changes with available inputs | Read what was measured before crediting one input for the full result. |
| Ranking and threshold decisions Source 1 · Published abstract and author manuscript | Matched participants and possible decision cutoffs | AUROC and threshold-specific error rates | A threshold for a new population requires separate validation. |
The export includes claims, available numbers, methods and source locations. It contains our reading notes and published summaries; it is not raw participant data or an independent reanalysis.
Evidence update · 22 Nov 2026
First publication. Any calculations in this brief use reported figures. The reading notes identify possible follow-up questions. AI source check; no human editorial or clinical review.
Model inputs and two different performance measures
| Inputs and model | Ranking measure | Probability error |
|---|---|---|
| Two voice tasks | AUROC 0.71 | Brier score 0.23. |
| Structured information only | AUROC 0.75 | Brier score 0.20. |
| All three inputs, full model | AUROC 0.83 | Brier score 0.18. |
| All three inputs, XGBoost comparator | AUROC 0.81 | Brier score 0.17. |
Author manuscript, Table 2. Values refer to the same matched-sample analysis.
Your questions, answered
Were the headline results obtained from voice alone?
The journal reports the combined model. The separate voice-only comparison in the author manuscript gives a lower AUROC of 0.71.
Does 0.83 AUROC mean the test got 83% of decisions right?
AUROC measures ranking across decision cutoffs. To read proportions of correct decisions, look for a stated cutoff and the associated test counts.
Why can a model rank better but have a higher Brier score?
A model can order cases well while the probabilities it assigns are still farther from the observed outcomes. Ranking and squared probability errors are different checks.
What does the half-and-half split mean?
The authors matched cases and controls for the study. Their equal numbers describe the research sample.
Limits of this interpretation
- The journal abstract supports the selected sample and headline results. Input comparisons are separately attributed to Table 2 of the author manuscript.
- The reported labels and balanced sample limit what the scores say about evaluation in a clinic.
- Changes in the collection setting, equipment, language or mix of respiratory symptoms could affect model performance.
Sources & transparency
- Despotovic et al. (2026) — Multimodal digital biomarker for asthma: complementary roles of vocal, clinical and demographic factors
Publisher’s abstract and publication record for the accepted article. Checked the sample size, self-reported reference label and headline performance. The author manuscript provides the detailed comparison. · Accessed 22 Nov 2026
DOI: 10.1038/s41746-026-03465-w - Despotovic et al. (2026) — Author manuscript on arXiv
Author manuscript, Tables 1 and 2 and Methods. The full-model values in Table 2 match the journal abstract. The earlier preprint summary lists different values, so detailed comparisons here identify their table. · Accessed 22 Nov 2026
DOI: 10.48550/arXiv.2607.08714
Prepared and source-checked with AI for Press-news Team. The source notes identify the material available during preparation. The article received no human editorial or clinical review. The source list supports the reported results and the calculations shown.
Source check: AI source check — primary research and stated source limits
Clinical review: Not applicable to this educational guide
Suggest a correction