PhD candidate, TU WienResearcher, Medical University of Vienna
Vienna, Austria
I work on offline reinforcement learning for clinical decision support in intensive care, and on evaluating when its recommendations can be trusted.
Research interests
I’m Mahdi, a machine-learning researcher working on offline reinforcement learning for critical care.
At the Medical University of Vienna, I develop and validate deep reinforcement learning methods for decision support in intensive care, and use off-policy evaluation to assess learned treatment policies on retrospective ICU data. I work with intensive-care physicians to check that the recommendations are clinically plausible.
I am also a PhD candidate in Computer Science at TU Wien. My background is in mechanical engineering, where I worked on state estimation and sensor fusion; the common thread is quantifying how reliable an estimate is.
From Kalman filtering to off-policy evaluation, an estimate is reported together with how far it could be wrong.
A treatment policy that looks best under one off-policy estimator may not under another, so I compare several before drawing conclusions.
Learned treatment strategies are reviewed with intensive-care physicians for clinical plausibility.
Machine learning for decisions under uncertainty.
My work spans state estimation from noisy sensors, reinforcement learning and language models, and clinical decision support. Across all three, the aim is to learn from imperfect data and to quantify how far the results can be trusted.
01
Kalman filtering and sensor fusion that turn low-cost, noisy sensor signals into reliable estimates of motion, together with their remaining uncertainty.
M.Sc. · Sharif University of Technology
02
Offline reinforcement learning, which learns decision policies from logged data rather than by trial and error, and off-policy evaluation of those policies; also hallucination detection in large language models.
PhD · TU Wien
03
Treatment strategies for sepsis and renal replacement therapy, learned from intensive-care records, evaluated with several off-policy methods and reviewed with physicians for clinical plausibility.
Medical University of Vienna
Five projects across three areas.
Objectives and outcomes in brief; the papers, code and data are linked from each project.
Learning fluid and vasopressor dosing policies for sepsis from intensive-care records (MIMIC-IV), and comparing off-policy evaluation methods, whose choice can change which policy appears best.
Conference abstract in Journal of Critical Care (ICCAI 2025, Vienna); full paper in preparation.
Learning when to start renal replacement therapy in critically ill patients with acute kidney injury, using reinforcement learning with high-confidence off-policy evaluation.
Conference abstract in Journal of Critical Care (ICCAI 2024); preprint on Research Square (2025).
Detecting hallucinations in language-model outputs (SemEval-2024 Task 6) with an ensemble that combines black-box detectors and detectors reading the model’s internal activations.
Ranked 3rd in the model-aware track and 8th in the model-agnostic track; paper in the SemEval-2024 proceedings (ACL).
Fusing drifting inertial sensors with camera-based SLAM in extended and square-root unscented Kalman filters to track a human arm, validated against an optical motion-capture system.
Journal article in Measurement (2023), with an open dataset and a graphical-abstract video (MP4).
Lead researcher for the Ludwig Boltzmann Institute for Digital Health and Patient Safety on a Horizon Europe project developing AI-enabled wearables for multisport training and remote healthcare.
Funded by Horizon Europe, grant agreement 101130495.
Papers, preprints & abstracts.
Six publications in machine learning for intensive care, language models and sensor fusion. Also listed on Google Scholar and ORCID.
Research Square Preprint
Journal of Critical Care 86, 154964 Conference abstract · ICCAI 2024
Iranian Journal of Engineering Education 24(94), pp. 39 to 60 Journal article · in Persian
Positions and education.