Hasindri Watawana

I am excited about the advances in NLP and speech understanding, and I look forward to finding ways to harness these breakthroughs for mental health. One thing I feel strongly about is the problematic trend of using opaque black-box models for mental health predictions — including generating evaluation rationales by asking the very same black box to explain itself, and then trusting those explanations. My PhD therefore focuses on building interpretable multimodal systems for mental health, currently focusing on SpeechLLMs, and I get to do this alongside the wonderful Dr. Esau Villatoro at IDIAP Research Institute & EPFL in Switzerland.

I graduated from University of Moratuwa, Sri Lanka, with a first class honours in Electronic and Telecommunication Engineering (BSc. Eng. Hons). My undergraduate thesis was titled Contrastive Deep Encoding Enables Uncertainty Aware Machine Learning Assisted Histopathology advised by Dr. Dushan Wadduwage, Dr. Ranga Rodrigo, and Dr. Chamira U. S. Edussooriya. I did a research internship with Dr. Kanchana Thilakarathna at University of Sydney during my undergraduate studies.

Prior to my PhD, I worked as a Research Assistant at MBZUAI, UAE, collaborating with Prof. Fahad Khan on foundation models for medical data and Large Multimodal Models (LMMs). Across all of these experiences, a common thread has always been using AI to benefit healthcare.

Email  /  Google Scholar  /  LinkedIn  /  Github

[Feb 2026]    When Consistency Becomes Bias: Interviewer Effects in Semi-Structured Clinical Interviews got accepted to LREC 2026 🎉.
[Jul 2025]    Joined IDIAP Research Institute & EPFL, Switzerland as a PhD Student — excited for the journey ahead! 🎉.
[Nov 2024]    ALM-Bench benchmark for evaluating LMMs on culturally diverse 100 languages has been released.
[Oct 2024]    Contrastive Deep Encoding Enables Uncertainty Aware Machine Learning Assisted Histopathology has been accepted to WACV 2025 🎉.
[Jun 2024]    HLSS has been accepted to MICCAI 2024 🎉.
[Mar 2024]    Hierarchical Text-to-Vision Self Supervised Alignment for Improved Histopathology Representation Learning (HLSS): preprint available.
[Jul 2023]    Joined MBZUAI, UAE as a Research Assistant.
[Jul 2023]    Contrastive Deep Encoding Enables Uncertainty Aware Machine Learning Assisted Histopathology: preprint available.
[Jul 2022]   Research presentation to Information Security and Privacy group of Data61, CSIRO, Australia [slides].
[Jan 2022]   Joined University of Sydney as a Research Intern.
[Aug 2021]   Our team DigitX emerged as champions in IEEE ICAS Student Challenge.

• Description: Investigated a systematic bias arising from interviewer prompts in semi-structured clinical interviews used for automatic depression detection. Models trained solely on interviewer turns — without any participant language — can match or outperform models trained on participant responses, across three datasets (DAIC-WOZ, ANDROIDS, E-DAIC) and two model architectures.


• Outcome: Demonstrated that this prompt-induced bias is cross-dataset and model-agnostic, and that restricting models to participant utterances leads to more distributed, genuine evidence. Findings highlight a key methodological concern for conversational mental health AI: reported performance gains may reflect scripted interview structure rather than true understanding of participant language.

• Description: Developed a novel language-tied histopathology image representation learning framework that explores the inherent hierarchy in histopathology image and text data.


• Outcome: Achieved state-of-the-art (SOTA) performance on two medical imaging benchmarks, OpenSRH and TCGA datasets. Our framework also provides better interpretability with the language aligned representation space.

• Description: Developed a self-supervised deep representation learning model for histopathology capable of assessing prediction uncertainty.


• Outcome: Achieved SOTA performance in patch and slide level classification on multiple cancer image datasets with only 1-10% annotations compared to benchmark. Our uncertainty-aware annotation method reaches SOTA with significantly fewer annotations compared to randomly selected annotation of data.

I borrowed this website layout from here!