AI in healthcare: today’s research roundup — 19 October 2026

Several new studies this week compare large language models directly against each other and against clinical benchmarks. In a blinded evaluation of 39 GERD-related questions, gastroenterologists rated Gemini-2.5 and Claude-4 above ChatGPT-5 on accuracy and comprehensiveness, while ChatGPT-5 produced the most readable, high-school-level responses, underscoring a clarity-versus-clinical-richness trade-off. A separate benchmark testing compact, on-device LLMs (gpt-oss-20b/120b) against GPT-5, o4-mini, and DeepSeek-R1 on diagnosis tasks found the smaller models performed comparably to larger cloud models, and fine-tuning closed most of the remaining gap to GPT-5. In psychiatric screening, a 555-interview benchmark found LLM accuracy ranging widely from 0.49 to 0.86 across anxiety, depression, and PTSD, with false negatives often tied to protective-context language overriding explicit symptom evidence in the models' own reasoning.

On the prognostic and diagnostic modeling side, a deep learning coronary-segmentation system reached Dice scores of 0.95 (heart) and 0.87 (LAD artery) across 283 externally validated patients and, in prospective use, flagged 19% of 1,386 radiotherapy patients for cardiology referral based on cardiac toxicity risk. A dermoscopic skin-lesion classifier using four architectures (ViT-B/16, Swin-S, ConvNeXt-S, EfficientNetV2-S) saw its ROC-AUC drop from 0.95-0.97 internally to 0.80-0.89 on external clinical validation, another reminder of the generalization gap that shows up repeatedly in this literature. Elsewhere, a wearable-only heart-rate model using a genetic-algorithm-tuned LSTM targeted early sepsis detection outside the hospital ward, and an XGBoost model paired with rule-based explanations predicted hospitalization need in COVID-19 outpatients from a 1,278-patient cohort.

Across both LLM and traditional ML studies, the same tension keeps surfacing: models that look strong in benchmarks or internal cohorts, and a persistent drop when tested externally or against real clinical nuance.

Sources

1. Do Bots Provide Correct and Adequate Guidance Regarding Acidity: A Blinded Comparison Rated by Patients and Physicians
2. Benchmarking and Adapting On-Device Large Language Models for Clinical Decision Support
3. When Symptoms Are Not Enough: Evidence-Weighting Patterns in Large Language Model Psychiatric Screening
4. Clinical Validation and Prospective Deployment of an Automated Deep Learning-Based Coronary Segmentation and Cardiac Toxicity Risk Prediction System
5. Cascade Classification of Dermoscopic Images of Skin Neoplasms with Controllable Sensitivity and External Clinical Validation
6. Early Prediction of Sepsis using Heart Rate Signals and Genetic Optimized LSTM Algorithm
7. Designing an Explainable Algorithm Based on XGBoost and Genetic Algorithm for Predicting Hospitalization Needs of COVID-19 Patients

Prepared with AI assistance from newly published research. See the sources for the original studies.