Bipin Lekhak

Bipin Lekhak

PhD candidate, Computing and Information Sciences
Rochester Institute of Technology · Computational Biomedicine Laboratory

I work on causal machine learning for time series. I build latent-variable models that learn a system's dynamics well enough to predict how it will respond to interventions: a tumour under chemo- or radiotherapy, a patient in intensive care, a physical system observed only through video. My models separate a system's native dynamics from its response to treatment, are identifiable, and adapt to each individual from its own history through meta-learning.

I am advised by Prof. Linwei Wang. Before the PhD I spent four years as a machine learning engineer at Fusemachines in Kathmandu, building medical image segmentation, 3D shape reconstruction and forecasting systems.

I am looking for postdoctoral and industry research positions starting in 2027 in causal inference, scientific machine learning and time-series modelling, for health and for engineered systems.

Portrait of Bipin Lekhak

News

Research

How will a system respond to an intervention it has not yet received? I build models that answer this from observational time series, and the guarantees needed to trust their answers.

Interventional latent dynamics

Forecasting a patient's trajectory under a candidate treatment plan needs two things at once: the patient's own disease dynamics, and how those dynamics change under treatment. ASIDE and its successor iASIDE model these as separate components of a latent ordinary differential equation. Meta-learned encoders read a subject's history and infer the parameters of each component, so the model adapts to every individual without retraining. iASIDE shows that the latent interventional dynamics are identifiable, and gives a separable formulation that reaches the identifiable solution in practice.

Schematic of the separable interventional dynamics model The observed history feeds three encoders: an initial-state encoder that outputs the latent state, a native meta-encoder that outputs native parameters, and a response meta-encoder that outputs response parameters. These drive a latent ODE whose derivative is the sum of a native term and a treatment-response term. Planned treatments enter the ODE, which a decoder turns into a forecast. Observed history x1 … xt a1 … at Initial-state encoder latent state zt Native meta-encoder parameters cn Response meta-encoder parameters ca Separable latent ODE ż = f(z; cn) + g(z, a; ca) native response integrated over the horizon planned treatments at+1 … Decoder forecast xt+1 …
The separable model. Encoders read the observed history and infer an initial latent state and the parameters of the native (blue) and response (orange) dynamics. The latent ODE is integrated forward under a planned sequence of treatments and decoded into a forecast.

Evaluated on tumour growth under chemo- and radiotherapy, MIMIC-III intensive-care data and videos of forced pendulums, with and without time-varying confounding, against established baselines such as the Causal Transformer, CRN and RMSN. The same structure applies to engineered systems driven by control inputs, where each unit behaves slightly differently and has to be modelled from its own history.

  • Scientific machine learning
  • Neural ODEs
  • Meta-learning
  • Identifiability
  • Variational inference
  • Time-varying confounding

Treatment effects over time

Separating native from response dynamics gives a natural handle on causal effects. With collaborators I am developing temporal versions of meta-learners, the X- and R-learners, that use this structure to estimate how outcomes would differ under alternative treatment sequences. This work also corrects for treatment-assignment bias with weighting and invariant representations.

  • Potential outcomes
  • Meta-learners
  • Inverse probability weighting

Remote monitoring after hospital discharge

Many heart-failure patients return to hospital within weeks of discharge. In an NIH-funded study with the University of Rochester Medical Center, I analysed data from a sensor-equipped toilet seat that recorded cardiac measurements at home for 90 days after discharge. We measured how many patients could be enrolled and how consistently they used the device, documented why in-home deployments failed, and showed that the seat's measurements predict hospitalisation within 21 days (AUC 0.74).

  • Remote patient monitoring
  • Care after discharge
  • Heart failure
  • Clinical risk prediction

Publications

Manuscripts under review

  1. Under review

    iASIDE: Identifiable, Adaptable, and Separable Interventional Dynamics

    Bipin Lekhak, Yubo Ye, Suraj Poudel, Ruby Shrestha, Linwei Wang

    A latent-dynamics framework that predicts how a system responds to hypothetical interventions. It establishes identifiability of the latent interventional dynamics by conditioning non-stationary parameters on history, and separates native from response dynamics so that optimisation can reach the identifiable solution.

Conference papers

  1. ICLR 2026

    ProReGen: Progressive Residual Generation under Attribute Correlations

    Ruby Shrestha, Ajay Gopi, Casey Meisenzahl, Bipin Lekhak, Linwei Wang

    International Conference on Learning Representations (ICLR), 2026

    OpenReviewPDF
  2. CinC 2023

    Novel In-Home Cardiac Monitoring for Heart Failure Patients

    Bipin Lekhak, Ryan Missel, Dillon Dzikowicz, Solomiya Rachynska, Wojciech Zareba, Linwei Wang

    Computing in Cardiology (CinC), 2023

    IEEE Xplore

Experience

Education