# Weekly AI Meeting ## An introduction — who I am, where I come from, and my research area *April 2026 · Manzoor Zarlish* --- ### About Me **Zarlish Manzoor** I'm a new PhD student from **Lahore, Pakistan**, joining the Department of Medical Informatics at Kyoto University as a D1 student. ### DPT Degree University of Sargodha, 2019–2024 ### Clinical Background Stroke rehabilitation, Gulab Devi Hospital ### Research Interest AI + Medical Informatics ## My Journey From Physical Therapy to Medical Informatics 1. **2019** Enrolled in DPT program at University of Sargodha 2. **2020–21** Ranked **1st** in First Professional Year; began exploring AI tools independently 3. **2022–24** Clinical rotations at Gulab Devi Hospital; six months in stroke rehab post-graduation 4. **2025** Explored research opportunities; connected with Prof. Kurado's lab 5. **April 2026** Joined Kyoto University, Department of Medical Informatics ## My Research Area Utilization of Artificial Intelligence in Medicine Affiliated with the Graduate School of Medicine ### Research Background A substantial subgroup of stroke rehabilitation patients shows minimal functional gains despite intensive therapy, and the variability in outcomes remains poorly understood. **Limited recovery** Some patients make only minimal progress even with intensive rehabilitation. **Short treatment window** The high-intensity rehabilitation period is only 2–4 weeks. **Early identification gap** Clinicians currently lack implementating of the tools due to trust **Research question:** Can we identify limited responders within the first few days of admission? ### Existing Literature More than 50 ML models for stroke rehabilitation outcomes have been published in the past two decades. A recent systematic review (Zu et al., 2023) included 10 studies comprising 32 prediction models, with reported AUC from 0.63 to 0.91. All were rated as having high or unclear risk of bias. Almost none of these models are used in routine clinical practice. Two reasons recur in the literature: 1. Most are black-box models. Clinicians do not act on a probability score they cannot interpret. 2. Most predict outcomes directly using arbitrary cut-offs, rather than recognizing the patient subgroups that actually exist in the data. ### The four phases 1. **Discover** — K-means++ clustering on FIM gain and FIM efficiency to surface patient phenotypes from the data 2. **Characterise** — Profile each phenotype using clinical, demographic, and downstream outcome variables 3. **Predict** — Multinomial logistic regression with LASSO regularisation to assign new patients to a phenotype within Day 1–3 4. **Explore** — Compare therapy utilisation patterns across phenotypes; generate hypotheses for future intervention trials ### Key design choices - Data-driven phenotypes rather than arbitrary thresholds - Interpretability treated as a primary design constraint, not a post-hoc add-on - SHAP for individual-level explanations at the bedside - TRIPOD+AI reporting standard throughout - Designed specifically for the Kaifukuki context, not retrofitted from a Western dataset - Approximately 300 to 350 historical stroke cases from collaborating Kaifukuki hospitals in the Kyoto region - Sample size derived using the Riley et al. (2020) prediction-model methodology rather than the legacy events-per-variable rule - Internal validation: bootstrap .632+ - External validation: temporal if data spans multiple years, otherwise geographic across collaborating hospitals - Missing data: Multiple Imputation by Chained Equations (MICE), with sensitivity analyses - Implementation in Python; code will be version-controlled and prepared for open release alongside publication - Methodological details. The choices in this proposal are my starting point. I expect to refine algorithm parameters, predictor sets, and validation specifics with the lab during protocol development. - Japanese clinical research culture. I want to learn the institutional norms of clinical informatics work in Japan, including how to engage with Kaifukuki hospitals and clinicians directly. - Japanese language. I am studying Japanese for both bedside engagement and lab life. - Publication and conference strategy. I would value the lab's view on which venues are the right fit for this kind of work, both in Japan and internationally. ### And one open question I would like to put to the room: - Within the lab's existing collaborations, which Kaifukuki hospitals would be the most natural starting point for the data access conversations? I am grateful for the chance to begin this work here, and I am looking forward to learning from this group. ### Discussion welcome. In particular, feedback on: - Methodological suggestions or concerns - Hospital and data partnership ideas - Conferences and journals the lab respects in this area - Anything I should be aware of as a new student in this lab and in Japan ## Goals ### What I Want to Learn · What I Hope to Contribute #### What I Want to Learn - Rigorous research methodology in medical informatics - Interpretable machine learning techniques for clinical data - How to design studies that clinicians actually trust and use - Japanese healthcare system and Kaifukuki data structures #### What I Hope to Contribute - A clinical perspective grounded in hands-on rehabilitation experience - Research that addresses real problems in stroke recovery prediction - Enthusiasm for cross-cultural, interdisciplinary collaboration - Work that could meaningfully support rehabilitation teams in Japan ## Research Direction ### Three Pillars of My Research My research aims to use **interpretable machine learning** to identify stroke patients who are likely to struggle in rehabilitation **.** This direction emerged from conversations with senior doctors during my clinical work, where the challenge of "poor responders" came up repeatedly. More detail will follow as the work develops. ### Thank You I'm genuinely grateful for the opportunity to join this lab and to learn from all of you. This is just the beginning — I'm looking forward to growing here, contributing where I can, and hopefully doing work that matters for patients and clinicians alike. ### Questions, thoughts, or suggestions are very welcome. I'd love to hear your perspective — especially from those who have worked with rehabilitation data or clinical ML before.