Research
My research currently spans two areas: AI for healthcare, particularly the use of language models to detect cognitive decline from speech, and AI safety and alignment, where I am studying unintended behavioral effects of safety-oriented training in large language models.
What Inspired My Research
My interest in Alzheimer’s research began partly from personal experience: my great-grandfather had Alzheimer’s disease. Growing up in Singapore, where an ageing population is making dementia an increasingly important healthcare challenge, made me interested in whether AI could help detect cognitive decline earlier and more accessibly. This led to my first research project on using speech-derived biomarkers and language models to detect Alzheimer’s disease.
My interest in AI safety and alignment developed later through my studies in philosophy. Questions about human values, moral reasoning, agency and responsibility made me increasingly interested in what happens when we try to translate human intentions into the behaviour of increasingly capable AI systems. That led me to my current research on emergent misalignment and alignment side effects, particularly whether safety-oriented training can create unintended behavioural changes elsewhere in a model.
Across both areas, I am interested in a broader question: how can we build AI systems that are not only capable, but reliable and aligned with human goals in domains where their decisions matter?
AI for Alzheimer’s Disease
Utilising Speech-derived Biomarkers to Detect Alzheimer's Disease with BERT-based Language Models: A Machine Learning Study
Zara Farida Khanna, Dean Ho, Alexandria Remus, Marlena Raczkowska, Railey Montalan and Pramit Saha
Frontiers in Artificial Intelligence, 2026
Accepted for publication
Abstract
Background: Early diagnosis is critical for effective management of Alzheimer's disease (AD). While prior studies have shown that speech features can be indicative of AD, most existing work consolidates multiple biomarkers, making it difficult to isolate the contribution of individual features.
Objective: This study systematically isolates individual speech biomarkers to quantify their distinct contributions to AD classification performance of language models (LMs) and determine whether targeted biomarker selection improves over consolidated feature sets.
Methods: We processed speech transcriptions from DementiaBank to surface discriminatory speech biomarkers—verbal pauses, disfluencies, and unintelligible words. We then fine-tuned and evaluated lightweight LMs (BERT, AlBERT, and DistilBERT) on these biomarker-conditioned transcripts for automatic AD classification.
Results: Pauses emerged as the most discriminatory speech biomarker (F1 = 0.8326), significantly outperforming the no-biomarker baseline and other biomarkers. Combining all biomarkers degraded performance relative to pauses alone on average across models, though the effect was model-dependent: BERT's all-biomarkers condition exceeded its own no-biomarker baseline, suggesting that feature combination benefits higher-capacity models. BERT yielded the best performance (F1 = 0.8154) across conditions.
Conclusions: Selective use of speech biomarkers such as pauses can meaningfully improve AD detection with lightweight LMs, suggesting that targeted biomarker selection may offer a more interpretable and clinically actionable path than broad feature consolidation.
AI Safety & Alignment
Emergent Misalignment and Alignment Side Effects in Coding-Focused Large Language Models
Research in progress
My current research examines whether interventions intended to make large language models safer can produce unintended behavioural changes elsewhere in the model.
Using coding-focused language models, I am studying the effects of alignment-oriented training on model behaviour, with particular interest in whether improvements in one safety-related objective can create unexpected trade-offs in others.
More broadly, the research explores a central question in AI alignment: when we intervene to make a model safer in one respect, what else changes?
Research Training
As part of my research at the National University of Singapore, I have completed research training through the CITI Program, Health Information Privacy and Security (HIPS) - Basic Course and Good Clinical Practice - Basic Course.
I have also completed university-level coursework relevant to my research in computer science, algorithms, machine learning and mathematics, including HarvardX’s CS50, Stanford’s Algorithms: Design and Analysis (Part I and Part II), Imperial College London’s Mathematics for Machine Learning Specialization, and Stanford’s Machine Learning Specialization.