🚨 Application deadline extended to July 18! . The ASGE AI Scholars Program offers a fully funded year of AI training for aspiring GI leaders. Get personalized mentorship, attend exclusive workshops, and contribute to a dynamic team project. Apply now:

RT @DOMSinaiNYC: Ali Soroush, MD, at Mount Sinai is leading efforts to apply generative AI in gastroenterology, from automating EHR data ex….

RT @AndrewYNg: I’d like to share a tip for getting more practice building with AI — that is, either using AI building blocks to build appli….

@ASGEendoscopy @sravmd @Ashwin_RaoV @tariqraseen @JamilSamaanMD @LakshmiSubbaraj @JosephHabibi_MD.

@AlyssaYChen14 is now on Twitter!.

Sa2015: CAN VISION LANGUAGE MODELS SEE WHAT THEY SAY? INVESTIGATING MODEL ATTENTION IN ULCERATIVE COLITIS SEVERITY SCORING AND COLORECTAL POLYP DETECTION. 3/.

Sa2075: ZERO-SHOT VISION-LANGUAGE MODELS COMPETE WITH TRAINED MACHINE-LEARNING MODELS ON POLYP DETECTION AND CLASSIFICATION. 2/.

Check out these cool AI posters showcasing our team's work today at #DDW2025!. Sa2035: MERGING INFLAMMATORY BOWEL DISEASE GUIDELINES USING LARGE LANGUAGE MODELS AND RETRIEVAL AUGMENTED GENERATION . 1/.

Thank you to:.- the the entire study team for their efforts: Suraj Jaladakani @pushkala2307 Ji Yoon Yoon Chin Hur @proftatonetti @girish_nadkarni. - @AGA_Gastro for their critical research support. 11/.

Limitations:.- Retrospective, single-center.- Missingness, errors in EHR data.- Small number of biopsy-confirmed events. We’re addressing this and more through:.📈 Additional modeling methods.🧠 LLM data extraction.🌍 External validation.👥 Algorithmic fairness metrics 10/.

🏥 Future clinical use case: .Risk-based triage in primary care or GI clinic . → Identify high-risk patients early with an automated risk calculator . → Identify and address modifiable risk factors . → Enables scalable, equitable screening programs 9/.

💡 Why this matters:.- Current guidelines often miss patients who don’t meet narrow subtype criteria—especially in non-White or immigrant groups. - Our approach aims to expand eligibility, improve equity, and support population-level risk stratification. 8/.

This is a proof-of-concept that:. ✅ Cumulative UGI cancer risk can be estimated from real-world data. ✅ Models can generalize across subtypes. ✅ Tools can be deployed without workflow disruption 7/.

⚙️ Model performance at optimal risk threshold (validation set):. AUROC: 0.81. Sensitivity: 80%. Specificity: 75%. NNS: 59 patients to detect 1 case 6/

After univariate analysis, we trained a Cox proportional hazards model using 13 structured features: FHx of gastric cancer, FHx of colon cancer, etoh use, race, ethnicity, GERD, peptic ulcer disease, Barrett’s esophagus, coronary artery disease, PPI use, and aspirin use. 5/

📍 Study design:.- 36M encounters from Mount Sinai Epic EHR.- Age 40–85.- Linked to cancer registry outcomes.- Excluded early, unconfirmed, or rare diagnoses.- Final cohort:. 👥 1.7M+ patients. 🔬 313 biopsy-confirmed UGI cancers 4/

🔍 What if we could estimate cumulative risk of any UGI cancer—before symptoms—using existing EHR data?.📊 No imaging.📝 No NLP.🙅 No manual review.Just variables available at the first visit. 3/

🧬 UGI cancers: EAC, ESCC, CGC, NCGC.🩸 ~4.5% of cancer deaths in the US.📉 5-year survival remains <30% due to late detection. Current screening is fragmented—based on subtype, not total risk. Many high-risk patients are missed. 2/

🧵 Can we unify upper GI cancer screening with cumulative risk prediction?. At #DDW2025, Alyssa Chen (PGY2!) presented our proof-of-concept risk model for predicting 5-year UGI cancer risk, using only readily accessible EHR data. Read more👇1/.

RT @NishantBalepur: 🚨 New Position Paper 🚨. Multiple choice evals for LLMs are simple and popular, but we know they are awful 😬. We complai….