chilconference

A multi-objective framework to fine-tune existing scoring tables (e.g., those found on mdcalc.com) to different cohorts and/or varying feature availability. Check out @keisenfong's poster at #CHIL2025

ML systems trained on treatment records often assume adherence. Using an LLM to extract adherence from clinical notes, authors show that ignoring non-adherence can reverse treatment effect estimates and harm model performance. arxiv.org/abs/2502.19625 #CHIL2025papers

🚑 New at #CHIL2025: We propose ExOSITO, an interpretable offline RL method for ICU lab test ordering. By leveraging side info + clinical rules, it reduces unnecessary tests while preserving critical care. #chil2025papers @jerryji2019 @rahulgk

Excited to highlight @WPotosnak et al.'s work: a novel hybrid global-local architecture + model-agnostic pharmacokinetic encoder that enables patient-specific treatment effect modeling—significantly improving blood glucose forecasting on large-scale datasets. #CHIL2025 @AutonLab

Introducing Time2Lang, a framework bridging Time-Series Foundation Models & LLMs for efficient health sensing beyond traditional text prompts. Check out how authors reprogram TFMs & LLMs for mental health! @ArvindPillai10, @spdimitris, @SkNepal #CHIL2025papers

New at CHIL! Authors propose a contrastive pretraining method for stress detection using multimodal data (wearables + surveys). Our CLIP-style framework boosts performance under limited labels on LifeSnaps & PMData. @YangZeyu9 @AkaneSano_ @RiceECE #CHIL2025papers

Introducing EHRXDiff: A novel framework for predicting future chest X-rays using prior image and medical events, dynamically tracking disease progression. Explore more: github.com/dek924/EHRXDiff @daeunkyung @junukim01 #CHIL2025papers

Using simulations & RWD, @SMukherjee89 shows ML-imputed phenotypes boost GWAS power only when built from upstream biomarkers. Downstream proxies inflate FDR, & high predictive R2 can mislead—genetic vs. environ. correlation matters. Pick proxies w/ causal insight! #CHIL2025papers

New at #CHIL2025: WatchSleepNet 💤 A novel, open-source model for smartwatch-based sleep staging using IBI signals. Outperforms SOTA with REM F1 = 0.63. Code & data below 👇 📂 github.com/willkewang/Wat… 📊 physionet.org/content/dreamt… @Big_Ideas_Lab #CHIL2025papers

🚨 Calling all health AI founders & builders! Join us at Health AI Builders: A CHIL Unconference — June 25th @ UC Berkeley. 💡 Real talk on AI, regulation, GTM, & fundraising 👥 Small-group convos, big impact 🎯 Apply to attend: lu.ma/2arsxv64 #CHIL2025 #HealthAI #ML4H

Introducing KEEP! A lightweight method that bridges knowledge graphs with real-world data to produce interpretable code embeddings; in our experiments, KEEP outperforms LM embeddings in semantic accuracy and clinical prediction tasks. @gamzeandgursoy #CHIL2025papers

🎉 Excited to kick off our #CHIL2025 research papers! Over the next few weeks, we’ll be highlighting the 42 cutting-edge accepted papers. Each one pushes the frontier of ML + health — stay tuned! 💡 #MachineLearning #HealthAI #ML4H