The latest interview in our series with the AAAI/SIGAI Doctoral Consortium participants features Thi Kieu Khanh Ho who is studying time-series anomaly detection. We found out more about her research, and what inspired her to study AI, and what she plans to work on next.
Tell us a bit about your PhD — where are you studying, and what is the topic of your research?
I am doing my PhD at McGill University and Mila – Québec AI Institute, in the Department of Electrical and Computer Engineering, supervised by Professor Narges Armanfard. My research focuses on time-series anomaly detection, the problem of teaching AI systems to recognize when something unusual or abnormal is happening in complex, real-world data streams, without relying on large amounts of labeled examples. The methods I develop are broadly applicable across domains, from healthcare to industrial monitoring, and predictive maintenance.
Could you give us an overview of the research you carried out during your PhD?
My PhD unfolded across several connected directions, with one thread running through all of it: how do you build AI models that detect anomalies reliably when labeled data is scarce, noisy, or contaminated?
Early on, I focused on self-supervised learning for anomaly detection – methods that learn meaningful representations from data without needing manually labeled anomalies, which are rare and expensive to obtain in practice. One key contribution was EEG-CGS, a contrastive and generative self-supervised framework that incorporates local graph structures to detect anomalous channels in EEG (electroencephalogram) recordings, with direct application to seizure detection. What I found exciting about this work is that it performs well without ever seeing a single labeled seizure during training.
I also developed TSAD-C, a graph- and diffusion-based framework for multivariate time-series anomaly detection that explicitly handles a tricky real-world problem: what happens when your training data itself is contaminated with anomalies? Most existing methods assume clean training data, which is rarely the case in practice, whether you are working with clinical signals, sensor data from industrial systems, or network traffic logs.
Alongside these, I contributed two surveys: one on self-supervised learning for anomaly detection, published in Neural Networks, and one on graph-based time-series anomaly detection, published in IEEE Transactions on Pattern Analysis and Machine Intelligence, which I think of as maps of the field: where we are, what the open problems are, and where the most promising directions lie.
More recently, I have been working on foundation models applied to physiological signals, specifically for identifying the epileptogenic zone in patients with drug-resistant epilepsy, in close collaboration with neurologists, clinicians, and AI teams from different institutions.
Is there an aspect of your research that has been particularly interesting?
The epileptogenic zone identification project is the one that stays with me the most. Drug-resistant epilepsy affects roughly a third of all epilepsy patients – people whose seizures cannot be controlled by any medication. For many of them, surgery is the only remaining option, and pinpointing exactly where in the brain seizures originate, from hours of intracranial EEG recordings, is extraordinarily difficult even for the most experienced clinicians.
Working on this problem meant sitting with the clinical reality of what the data represents: real patients who have exhausted every available treatment. That changes how you think about model performance entirely. It is not about beating a benchmark; it is about whether your system is trustworthy enough to inform a surgical decision. That responsibility pushed me to think much more carefully about uncertainty, explainability, and failure modes than I ever had before, and convinced me that the gap between a good research result and a genuinely useful clinical tool is much larger than it appears on paper.
What are your plans for building on your research — what aspects will you be investigating next?
There are two directions I am most excited about. The first is pushing anomaly detection methods toward real-world deployability, moving beyond controlled benchmark evaluations toward settings where the data is messier, anomaly types are richer, and deployment constraints are real. My work on contaminated training data and open-set anomaly detection is a step in that direction, but there is much more to explore.
The second is continuing to develop foundation models for physiological and time-series data more broadly. Large-scale pretraining has transformed computer vision and natural language processing (NLP), I believe the same shift is coming for physiological and clinical AI, and my epileptogenic zone work is an early example of what that can look like. The key challenge is building models that are not just accurate, but interpretable and robust enough that practitioners, whether clinicians, engineers, or operators, can actually trust and use them in the real world.
What made you want to study AI?
Honestly, what drew me in was the universality of it. Early in my studies, I was working on biosignal processing, trying to classify brain activity, detect anomalies in physiological signals, and I kept noticing that the same mathematical tools that worked for one problem translated surprisingly well to completely different ones. There was something almost unreasonably powerful about that generality. Once I saw it, I could not unsee it.
What kept me in the field, though, was the combination of intellectual depth and real-world stakes. Anomaly detection sounds abstract, but the applications are deeply concrete, finding the brain region causing a patient’s seizures, catching a fault in an industrial system before it becomes a disaster, detecting a cyberattack before it causes damage. That breadth is what I love most about the field. Every new application domain teaches you something new about what your methods can and cannot do, and that tension between generality and specificity is endlessly interesting to work through.
Could you tell us an interesting (non-AI related) fact about you?
I have a genuine passion for singing; I came second in a K-pop singing contest at my university, which I consider one of my more underrated achievements, and music has been a constant throughout my life no matter where I have lived or studied. I am also an enthusiastic photographer; there is something about slowing down, looking carefully, and noticing details that most people walk past that gives my brain a useful reset from research. Outside of that, I love exploring new cultures, learning languages, and finding unexpected connections between things that seem completely unrelated at first glance, which, come to think of it, is also a fairly good description of what anomaly detection is about.
About Thi Kieu Khanh Ho
Thi Kieu Khanh Ho is a PhD candidate in Electrical and Computer Engineering at McGill University and Mila – Québec AI Institute, supervised by Professor Narges Armanfard. Her research focuses on self-supervised learning, graph-based methods, foundation models, and anomaly detection for time-series data, with applications spanning healthcare, industrial monitoring, and safety-critical systems. She is a Vanier Canada Graduate Scholar, FRQNT Scholar, and recipient of the McGill Engineering Doctoral Award. Her work has been published at AAAI, UAI, WACV, ECAI, IEEE TPAMI, and Neural Networks, with multiple papers selected for oral presentation. She co-organizes the workshop on automated spatial and temporal anomaly detection and tutorials on time-series anomaly detection at top-tier AI venues.