Dr. Newton Mwai
PhD in AI & Machine Learning
Email: nkinyanj[at]alumni.cmu.edu
Interests: Machine Learning, Interactive Personalization, Recommender Systems, Experimentation, Bandit Algorithms, Causality, Machine Learning for Health, Trustworthy AI
About Me
I attained my PhD in Computer Science and Engineering in the Healthy AI Lab at Chalmers University of Technology and the Wallenberg AI, Autonomous Systems and Software Program (WASP) Sweden. My research focused on designing interactive sequential learning algorithms for personalization, particularly in healthcare settings. I developed sample-efficient bandit algorithms that leverage historical data and incorporate real-world constraints to improve treatment personalization. This work is also applicable to broader interactive personalization tasks, such as content recommendation, adaptive experimentation (A/B testing), and off-policy evaluation.
Before my PhD, I attained a Master's degree in Electrical and Computer Engineering at Carnegie Mellon University and dual Bachelor's degrees in Electrical and Electronics Engineering and Telecommunication and Information Engineering at Dedan Kimathi University of Technology. I have also worked as a software and infrastructure engineer at Adtel (Jamii Telecom), interned at IBM Research, and contributed off-policy evaluation estimators to Vowpal Wabbit through Microsoft Research's RL Open Source Fest.
PhD Thesis
Leveraging Structural Priors and Historical Data for Practical Treatment Personalization with Multi-Armed Bandits. Defended on October 10th 2025 at Chalmers University of Technology. [Thesis]
Efficient and Practical Treatment Decisions with AI
Finding the right treatment for a patient often involves testing several options before settling on the most effective one. In chronic diseases such as Alzheimer’s, where treatment benefits are short‑lived and regular adjustments are needed, this trial‑and‑error process can be a long and costly process for both patients and healthcare systems. The aim of this research is to develop AI-based methods to shorten this search, while also respecting practical constraints, for example, reducing how often treatments are switched to avoid discomfort and side effects.
The work adapts a family of algorithms known as multi‑armed bandits, designed for learning through trial and error, for application in medical decision-making. It introduces new strategies that use patterns hidden in past records of patient treatment to guide quicker, effective decisions for new patients, explores ways to model patient differences flexibly, and develops methods that limit treatment changes without harming effectiveness. To test these algorithms safely, a realistic simulation environment for Alzheimer’s treatment was built using real clinical data and expert medical knowledge.
The thesis paves the way for more efficient and patient‑friendly decision‑making with AI in personalized medicine for the future.
På svenska: Effektiva och praktiska behandlingsbeslut med AI
Att hitta rätt behandling för en patient innebär ofta att man måste prova flera alternativ innan man hittar det mest effektiva. Vid kroniska sjukdomar som Alzheimers, där behandlingseffekterna är kortvariga och regelbundna justeringar behövs, kan denna process av att pröva sig fram bli både lång och kostsam för patienter och vårdsystem. Målet med den här forskningen är att utveckla AI-baserade metoder som förkortar sökandet, samtidigt som praktiska begränsningar respekteras – till exempel att minska hur ofta behandlingar byts för att undvika obehag och biverkningar.
I avhandlingen används en grupp av algoritmer som kallas multi-armed bandits vilka är framtagna för lärande genom trial and error, för användning i medicinskt beslutsfattande. De introducerar nya strategier som hittar och använder mönster från tidigare patientjournaler för att kunna fatta snabbare och mer effektiva beslut för nya patienter. Vidare undersöks flexibla sätt att modellera patienters olikheter för att utveckla metoder som begränsar behandlingsbyten utan att försämra effektiviteten. För att testa algoritmerna på ett säkert sätt byggdes en realistisk simuleringsmiljö för Alzheimerbehandling, baserad på verklig klinisk data och medicinsk expertkunskap.
Avhandlingen banar väg för mer effektiva och patientvänliga beslut med AI inom framtidens individanpassade medicin.
Publications
- 2026: Ahmet Zahid Balcıoğlu, Newton Mwai, Emil Carlsson, and Fredrik D Johansson. Identifiable Latent Bandits: Leveraging observational data for personalized decision-making. Transactions on Machine Learning Research, 2026. [Paper]
- 2025: Newton Mwai, Emil Carlsson, and Fredrik D Johansson. Latent Preference Bandits. Preprint, 2025. [Paper]
- 2025: Newton Mwai, Milad Malekipirbazari, and Fredrik D Johansson. Understanding exploration in bandits with switching constraints: A batched approach in fixed-confidence pure exploration. European Workshop on Reinforcement Learning (EWRL), 2025 [PDF]
- 2025: Lena Stempfle, Anton Matsson, Newton Mwai, and Fredrik D Johansson. Prediction models that learn to avoid missing values. In International Conference on Machine Learning (ICML), 2025. [Paper]
- 2025: Christoffer Ivarsson Orrelid, Oscar Rosberg, Sophia Weiner, Fredrik D Johansson, Johan Gobom, Henrik Zetterberg, Newton Mwai, and Lena Stempfle. Applying machine learning to high-dimensional proteomics datasets for the identification of Alzheimer's disease biomarkers. Fluids and Barriers of the CNS, volume 22, page 23. Springer, 2025. [Paper]
- 2024: Newton Mwai, Milad Malekipirbazari, and Fredrik D Johansson. Batched fixed-confidence pure exploration for bandits with switching constraints. In ICML 2024 workshop: aligning reinforcement learning experimentalists and theorists, 2024. [Paper]
- 2023: Newton Mwai, Emil Carlsson, and Fredrik D Johansson. Fast treatment personalization with latent bandits in fixed-confidence pure exploration. Transactions on Machine Learning Research, 2023. [Paper] [Code]
- 2023: Newton Mwai. Improved Sequential Decision-Making with Structural Priors: Enhanced Treatment Personalization with Historical Data. Licentiate thesis, Chalmers University of Technology, Sweden, 2023. [Paper]
- 2022: Newton Mwai and Fredrik D Johansson. ADCB: An Alzheimer's disease simulator for benchmarking observational estimators of causal effects. In Conference on Health, Inference, and Learning, pages 103-118. PMLR, 2022. [Paper] [Code]
- 2020: Newton Mwai, Timothy Odonga, Celia Cintas, Noel CF Codella, Rameswar Panda, Prasanna Sattigeri, and Kush R Varshney. Fairness of classifiers across skin tones in dermatology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 320-329. Springer, 2020. [Paper]
- 2019: Newton Mwai, Timothy Odonga, Celia Cintas, Noel CF Codella, Rameswar Panda, Prasanna Sattigeri, and Kush R Varshney. Estimating skin tone and effects on classification performance in dermatology datasets. In NeurIPS Workshop on Fair ML for Health, 2019. [Paper]