Monthly update - April 2025

First I would like to congratulate my colleagues and friends Jakob German and Mattia Cordioli for their publication titled "Association between plausible genetic factors and weight loss from GLP1-RA and bariatric surgery" in nature medcine!

Also congratulations to Feiyi Wang and Aoxing Liu for their publication "Effects of parental autoimmune diseases on type 1 diabetes in offspring can be partially explained by HLA and non-HLA polymorphisms" in Cell genomics!

Top Picks of the Month

This month featured several interesting publications from many different fields, which made choosing a standout particularly difficult. I think I’ll give the crown to the Google Research team and their two Nature papers exploring the use of large language models in clinical diagnosis and patient interaction.

In the first paper, they introduce AMIE—a large language model designed for diagnostic reasoning—and evaluate its ability to generate differential diagnoses both independently and as a support tool for clinicians.

Towards Accurate Differential Diagnosis with Large Language Models

The second study examines AMIE’s broader performance across various aspects of clinical interaction, including history-taking, management, communication, and empathy. The model was tested against primary care physicians in a randomized, double-blind crossover study.

Towards Conversational Diagnostic Artificial Intelligence

Other Notable Mentions

A Multi-Sequence MRI-Based Hierarchical Expert Diagnostic Method for the Molecular Subtype of Breast Cancer

• DATA: proprietary

• METHOD: Hierarchical Mixture of Experts (HMoE)

Continuous sleep depth index annotation with deep learning yields novel digital biomarkers for sleep health

• DATA: SHHS, CFS, MESA, MROS

• METHOD: Transformer encoder + MLP decoder, Gaussian mixture model

Development and validation of multimodal deep learning algorithms for detecting pulmonary hypertension

• DATA: Shenzhen Fuwai Hospital, Fuwai Yunnan Cardiovascular Hospital

• METHOD: multi-modal fusion model (DAE + BERT + ViT)

Conformal prediction enables disease course prediction and allows individualized diagnostic uncertainty in multiple sclerosis

• DATA: Swedish MS Registry (SMSReg)

• METHOD: LSTM + conformal prediction

Implementation of a national AI technology program on cardiovascular outcomes and the health system

• DATA: FISH&CHIPS cohort

• METHOD: Fractional Flow Reserve-Computed Tomography (FFR-CT)

Meta-prediction of coronary artery disease risk

• DATA: UKBioBank, AllofUs

• METHOD: meta-prediction (XGBoost of meta-features and baseline features)

Noninvasive early prediction of preeclampsia in pregnancy using retinal vascular features

• DATA: Sun Yat-sen University First Affiliated Hospital (FAH) cohort

• METHOD: multi-branch U-Net (RMHAS) for segmentation + LASSO logistic regression for prediction

Vision transformer based interpretable metabolic syndrome classification using retinal Images

• DATA: H-PEACE study , ImageNet-1K, MEH-MIDAS, Kaggle EyePACS

• METHOD: ViT pretrained with MAE

Assessment of a Polygenic Risk Score in Screening for Prostate Cancer

• DATA: BARCODE1 study

• METHOD: PRS

Common-variant and rare-variant genetic architecture of heart failure across the allele-frequency spectrum

• DATA: 10 biobanks

• METHOD: -

LEOPARD. missing view completion for multi-timepoint omics data via representation disentanglement and temporal knowledge transfer

• DATA: MGH COVID study, KORA study

• METHOD: conditional generative adversarial nets (cGAN)

The circulating proteome and cancer risk. A systematic literature review and meta-analysis of 26 prospective studies with genetic validation

• DATA: Embase and Medline

• METHOD: systematic review & meta-analysis

New Threshold for Defining Mild Aortic Stenosis Derived From Velocity-Encoded MRI in 60,000 Individuals

• DATA: UKBB

• METHOD: AU-Net

Adherence to life’s essential 8 is associated with delayed white matter aging

• DATA: UKBB

• METHOD: random forest + linear regression

Premature Mortality Attributable to Ultraprocessed Food Consumption in 8 Countries

• DATA: 8 country cohorts

• METHOD: meta-analysis

Application of large language models in medicine

• DATA: -

• METHOD: LLM

Automated Transformation of Unstructured Cardiovascular Diagnostic Reports into Structured Datasets Using Sequentially Deployed Large Language Models

• DATA: Yale New Haven Health System, MIMIC III, MIMIC IV

• METHOD: LLM

Benchmarking large language models for biomedical natural language processing applications and recommendations

• DATA: 12 BioNLP benchmarks

• METHOD: LLM

Foundation models for electronic health records. representation dynamics and transferability

• DATA: MIMIC-IV + University of Chicago Medical Center (UCMC)

• METHOD: foundation model

High-precision information retrieval for rapid clinical guideline updates

• DATA: 1 guideline database + 3 clinical trial databases

• METHOD: Named Entity Recognition (NER) and Normalization (NEN) + extraction of Population--Intervention--Comparison--Outcome (PICO) spans.

Retrieval augmented generation for 10 large language models and its generalizability in assessing medical fitness

• DATA: 35 singapore & 23 international guidelines + UpToDate articles

• METHOD: LLM + Auto-Merging RAG

Sociodemographic biases in medical decision making by large language models

• DATA: Mount Sinai Health System

• METHOD: LLM

Final Notes

HAI released the "2025 AI Index Report", chapter 5 is about "Science and Medicine", read more about it here: https://hai.stanford.edu/ai-index/2025-ai-index-report/science-and-medicine

This month I scanned a total of 30 publications spanning 12 different countries.

My top picks are based on my topics of interest—definitely a bit biased! If you came across any cool papers published in April 2025 (related to health data science topics) that you think deserved more recognition, please send them my way and I’ll try to include them in next month’s update!