Shishir Rao

Artificial intelligence for electronic health records For our next talk, in Machine Learning, we will be hearing from Dr Shishir Rao, Senior Research Scientist, Deep Medicine research group, NDWRH, University of Oxford. We’re delighted to host Shishir in...

Looking forward to presenting our team's research on AI for EHR at The Big Data Institute at Oxford on July 10th at 2 pm! Thanks to Sumeeta and Prof @nichols.bsky.social for organising.

Link: talks.ox.ac.uk/talks/id/590...

04.07.2025 12:38 — 👍 0    🔁 1    💬 0    📌 0

Refined selection of individuals for preventive cardiovascular disease treatment with a transformer-based risk model TRisk enabled a more targeted selection of individuals at risk of cardiovascular disease in both the primary prevention population and cohorts with diabetes, compared with benchmark approaches. Incorp...

This could not have transpired without the leadership of Prof Kazem Rahimi and team: Drs/Profs. Yikuan Li, Mo Mamouei, Milad Nazarzadeh, @rezakhorshidi.bsky.social, @gscollins.bsky.social, @vickersbiostats.bsky.social, Gosia Wamil, Chris Yau, Rod Jackson, & Goodarz Danaei!

🔗: tinyurl.com/LDHTRisk

12.06.2025 10:12 — 👍 1    🔁 0    💬 0    📌 0

🌍 Why it matters:
TRisk offers a scalable, equitable approach to prevention. More than just achieving strong predictive performance, TRisk can reduce over-treatment - crucial for efficient allocation of therapies - while identifying those most likely to benefit⚕️🫀🤖

12.06.2025 10:12 — 👍 0    🔁 0    💬 1    📌 0

In primary prevention cohort, TRisk at 15% threshold reduced patients selected for preventive therapy by 34% without gain in false negatives with respect to status-quo approach. In diabetes cohort, similarly TRisk at 10% reduced patients selected by 25% and 15% with minimal false negatives with respect to status-quo "treat all" and QRISK3 approaches respectively.

📉 What this means for CVD prevention:
In the general population:
🩺 TRisk reduced treatment recommendations by 34% vs QRISK3, without increasing false negatives.

In diabetes:
🩺 24% fewer treatment recommendations w respect to "treat all" strategy, with minimal missed cases (0.2%).

12.06.2025 10:12 — 👍 0    🔁 0    💬 1    📌 0

⚙️ What we found:
For incident CVD prediction, TRisk showed:
✅ Higher discrimination (C-index 0.91) and net benefit than benchmark models
✅ Better performance across age, sex & socio-economic deprivation groups
✅ Less sensitivity to age ranges
✅ Excellent calibration

12.06.2025 10:12 — 👍 0    🔁 0    💬 1    📌 0

TRisk architecture uses Transformer embeddings and attention built on the BEHRT model. Ordinary differential equations based neural network modelling framework is used for survival modelling.

🧠 How we built it:
TRisk is a Transformer-based survival model trained on:
🔹 2.2M patient records for training
🔹 750K for validation
🔹 3.8k diagnoses, 390 meds, 1.4k tests for model input

Importantly, it uses raw, longitudinal EHR sequences (figure) - no manual variable selection and no imputation.

12.06.2025 10:12 — 👍 0    🔁 0    💬 1    📌 0

📚 Why now?
Current tools for CVD prediction like QRISK3 often over-recommend treatment—especially in diabetes, where everyone is treated regardless of risk.

TRisk changes that. It analyses full EHR histories to predict CVD risk more precisely, not just based on age or a fixed checklist.

12.06.2025 10:12 — 👍 0    🔁 0    💬 1    📌 0

Thanks @erictopol.bsky.social🙌 Kazem and I are excited to share our paper, now in press at The Lancet Digital Health! We introduce TRisk, a Transformer-based survival model for 10-year CVD risk prediction and more refined preventive therapy recommendations 🫀📊

A short 🧵...

12.06.2025 10:12 — 👍 4    🔁 0    💬 1    📌 0

Appreciate the shout-out @erictopol.bsky.social!

11.06.2025 01:15 — 👍 1    🔁 0    💬 0    📌 0