Winston Chen

Healthcare & RL & Causality | PhDing @michigan_AI | Via @uw_ece @uwgenome @nvidia @realnetworks

[Winston Chen @winstonchenn]

@Aishwarya_R_M @BeEngelhardt @EmmaBrunskill @lschmidt3 @Emily_Alsentzer Congrats! Looking forward to your future healthcare work!

[Shengpu Tang @shengpu_tang]

Featured in today's STAT "AI Prognosis" newsletter: Your sepsis algorithm shouldn't require a time machine 🕗

Check it out for an explainer of what we found in our recent npj Digital Medicine article.

[Winston Chen @winstonchenn]

📄Measuring Model Performance in the Presence of an Intervention (AAAI’26, Special Track on AI for Social Impact):
- Paper: https://t.co/HGvu4Y23D8
- Code: https://t.co/epvejtPKQo
I will be present this paper Thursday (1/22) in Hall 2, board 34 between 12-2PM. Happy to chat about model/policy evaluation under interventions & efficient use of RCT data.

[Winston Chen @winstonchenn]

🚨 Model evaluation gets tricky when outcomes are affected by an intervention, even if the intervention is not driven by the model’s predictions.

In our #AAAI2026 paper, we study this issue in AI for social impact settings, where outcomes differ under intervention vs no intervention, and you need to be precise about what you’re evaluating. 🧵⬇️

[Winston Chen @winstonchenn]

Takeaway:
- Researchers evaluating models with limited RCT data should consider using NPW to improve sample efficiency.
- NPW provides a principled way to boost sample efficiency without introducing bias.

[Winston Chen @winstonchenn]

If you’re interested in AI for scientific discovery, check out our work on error-controlled interaction discovery, which is published on NMI today!

[Winston Chen @winstonchenn]

@thegautamkamath @taha_moji Sounds like a tradeoff, spending less time on each paper also reduces the probability for each paper to be well received.

[Winston Chen @winstonchenn]

See a blurb on this work here

[chilconference @CHILconference]

📌Tired of ignoring non-adherence in intention-to-treat analysis? Estimate assignment effects more efficiently with the front-door adjustment.

Meet LobsterNet🦞— a front-door method outperforming backdoor, especially when effects are small!

#CHIL2025papers

[Winston Chen @winstonchenn]

I’ll present my work on causal effect estimation under treatment non-adherence this week at #CHIL2025 looking forward to connect with everyone in Berkeley!

[MichiganAI @michigan_AI]

Excited to unveil this new research by @winstonchenn, @chang_trenton, & Jenna Wiens, check it out at #CHIL2025!

📄"Conditional Front-door Adjustment for Heterogeneous Treatment Assignment Effect Estimation Under Non-adherence"
🧠https://t.co/4RiZ6RWpXy

[MichiganAI @michigan_AI]

🚀New research represents a significant step towards developing graph positional encodings that effectively capture complex structures in heterophilous graphs. #AISTATS2025

Work by Michael Ito, Jiong Zhu, Dexiong Chen, Danai Koutra, and Jenna Wiens:
📄https://t.co/TzNZmSRYk0

[Winston Chen @winstonchenn]

@joejanizek Thanks for the post, Joseph! I’m glad you found it a cool idea!