HRDPath Scores Higher on DNA-Repair Deficiency Classification

Evidence

HRDPath performed numerically better than its comparator at identifying a tumor DNA-repair deficiency from ovarian-cancer tissue images. Its target, homologous recombination deficiency, is impairment of a specific DNA-repair pathway.

The HRDPath study developed the model using stained whole-slide tissue images without first outlining tumor regions. Its development cohort included 152 patients: 145 with high-grade serous ovarian or fallopian-tube cancer and seven with other ovarian cancers. Researchers also used two independent, previously unseen validation datasets.

On the reported test dataset, HRDPath’s area under the receiver operating characteristic curve—a measure of how well a model separates outcome groups across classification thresholds—was 0.846 versus 0.808 for the comparator, DSMIL. Sensitivity, the proportion of deficient cases correctly identified, was 69.2% versus 53.8%. Specificity, the proportion of non-deficient cases correctly identified, was 93.8% versus 87.5%.

The diabetic sepsis study retrospectively analyzed records from 7,716 adults with diabetes and sepsis—a life-threatening response to infection—in intensive care at one hospital during 2008–2022. In-hospital deaths numbered 1,357, or 17.6%. Researchers compared six predictive models with the Sequential Organ Failure Assessment score and randomly divided records into training and internal-validation groups.

Gradient Boosting Decision Tree, a model combining decision trees, had the strongest reported internal discrimination. Its curve-area score was 0.792, with a 95% confidence interval of 0.769–0.814 expressing statistical uncertainty. Validation sensitivity was 76.7% and specificity 68.6%. Precision was 34.2%: among patients the model classified as deaths, 34.2% actually died in hospital.

The Austrian hospital analysis combined annual cross-sectional assessments—snapshots rather than individual follow-up—from 2017–2025. It included 9,991 hospital patients aged 65 or older. The Malnutrition Universal Screening Tool identified nutritional risk in 23.4% with documented dementia versus 23.7% without, with no statistically detectable difference reported. Among patients at risk, mealtime assistance was documented in 59.7% with dementia versus 15.3% without; oral nutritional supplements were documented in 28.5% versus 19.2%.

The dermatology consensus addresses Janus kinase inhibitors—drugs targeting enzymes involved in immune signaling—across inflammatory diseases, including atopic dermatitis, alopecia areata and psoriasis. Two researchers independently evaluated evidence, resolving disagreements through discussion or a third expert. This was an expert-consensus process, not a skin-outcome study.

Analysis

Interpretation: The two modeling studies share a measurement approach, not a biological finding. Both assess discrimination: separating patients or tumors into outcome groups. Their scores cannot fairly rank the models against each other because the populations, targets and validation designs differ.

The nutrition findings distinguish screening prevalence from care documentation. More frequently documented assistance or supplements do not establish that those interventions caused better outcomes. Likewise, stronger model scores do not demonstrate that using a model improves care.

Limitations

HRDPath’s quoted scores are reported test-dataset results; they should not be assigned to either independent validation cohort. The briefing provides no confidence intervals for the model comparison, so numerical superiority does not establish statistically reliable superiority.

The sepsis model underwent only single-center internal validation. Its performance may change in other hospitals, and retrospective records can introduce selection and information bias.

The nutritional assessments cannot establish causality. Similar observed screening proportions do not prove equivalence, and documented interventions do not establish their appropriateness or benefit. Changes in participating hospitals and patients also limit interpretation of patterns over time.

The dermatology briefing provides no skin-specific comparative effect estimates. Its consensus methods therefore cannot quantify benefits or harms for individual skin diseases.

What to watch

Prospective evaluation across multiple hospitals would test whether the sepsis model retains its precision and whether predicted mortality matches observed mortality. For HRDPath, a confidence-changing result would be a prespecified comparison with DSMIL in an additional external cohort, reporting confidence intervals for differences in discrimination, sensitivity and specificity.