Truace
Subscribe

Urinary extracellular vesicle microRNA profiles and prostate cancer biopsy outcome: Moderate discrimination and quantified optimism under leakage-controlled machine learning

Background: Urinary extracellular-vesicle-associated microRNAs are candidates for non-invasive prostate cancer assessment. We evaluated discrimination between GSE159177 records annotated with a Gleason Score and those labelled Control; individual negative-biopsy status could not be confirmed. Results: The frozen analysis set comprised 465 urinary sample records (278 Gleason-labelled; 187 Control-labelled). Biological annotation retained 2578 mature human microRNA probes from 6658 array probes. Five-fold nested c…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1764· cite & verify
Urinary extracellular vesicle microRNA profiles and prostate cancer biopsy outcome: Moderate discrimination and quantified optimism under leakage-controlled machine learning
Laboratory Test Tubes; from a medical laboratory in Abuja, Nigeria by Frankincense Diala. CC0 · http://creativecommons.org/publicdomain/zero/1.0/deed.en

In brief

Researchers used machine learning on a public dataset (GSE159177) of 465 urinary sample records, 278 Gleason-labelled and 187 Control-labelled, to see whether extracellular-vesicle microRNA profiles could separate them. Feature selection, scaling, model choice and thresholds were confined to training data in five-fold nested cross-validation. Mean AUC across five repetitions was 0.707.

The work matters because it quantifies how much common analysis shortcuts inflate apparent performance for a non-invasive prostate cancer test. Even so, sensitivity was limited and individual participant independence could not be confirmed, so the findings remain exploratory and need replication at participant level.

Main points

  1. The analysis set had 465 urinary sample records, with 2578 mature human microRNA probes retained from 6658 array probes.
  2. Mean pooled AUC was 0.707 over five repetitions, with a range of 0.681 to 0.726.
  3. The original seed-42 result of AUC 0.726 was the highest of the repetitions, and its thresholded predictions gave 55.8% sensitivity.
  4. Selecting features with global outcome information raised AUC by 0.041, quantifying one source of optimism.
  5. hsa-let-7b-5p was the only microRNA significant after Benjamini-Hochberg correction, and it remains exploratory.

The problem

With leakage-controlled nested cross-validation, the model reached only moderate discrimination and missed 44.2% of positive-labelled records, so it is not suitable for clinical use.

The rundown

The study compared Gleason-labelled records with Control-labelled records, but individual negative-biopsy status could not be confirmed. The comparison therefore does not directly measure biopsy outcomes for individual participants.

A matched experiment showed that choosing features using the outcomes across the full dataset inflated AUC by 0.041 (paired 95% confidence interval 0.003-0.080). This gives an empirical estimate of the optimism that leakage introduces.

The authors state that repetitions reused the same records and were not independent validation. They call for targeted experimental confirmation and participant-level validation before clinical translation, and note that clinical value and economic benefits were not evaluated.

Sources

  1. Peer-reviewedBiochemistry and Biophysics Reports2026-10-01

The debate