research
Thirty Scaled Features Match the Deep Respiratory Model
In patient-grouped evaluation, standardized SVM features reached 0.78 macro recall without a 449-megabyte deep-learning runtime.
Summary
In patient-grouped evaluation, standardized SVM features reached 0.78 macro recall without a 449-megabyte deep-learning runtime.
On 1,211 respiratory clips from 201 patient groups, standardizing 30 audio-summary features moved an SVM from 0.51 to 0.78 macro recall—a 27-point effect larger than the measured architecture gap. Deep and classical models were not statistically separated, while clip-level splits inflated recall by 7.5 to 9.8 points. The dataset also carries recording-source confounding, which the authors identify as a main limitation.
Why it matters
In patient-grouped evaluation, standardized SVM features reached 0.78 macro recall without a 449-megabyte deep-learning runtime.
Limits and context
- Deep and classical models were not statistically separated, while clip-level splits inflated recall by 7.5 to 9.8 points.
Key claims
In patient-grouped evaluation, standardized SVM features reached 0.78 macro recall without a 449-megabyte deep-learning runtime.
Qualification: Deep and classical models were not statistically separated, while clip-level splits inflated recall by 7.5 to 9.8 points.
Evidence: source-2026-10-27-017
Sources
- arXiv preprint 2610.33030, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.