Ten years ago, deep learning was not the widely recognised phrase it is today. It was, as Professor Truyen Tran recalls, “just another subject mentioned in the curriculum”; a niche corner of computer science that could recognise a face or translate a sentence or two, but little more. The idea that a machine could read a decade of someone’s medical history and forecast where their health was heading sounded, at best, wildly ambitious.
That was the bet a young PhD student named Trang Pham made at Deakin’s then-Centre for Pattern Recognition and Data Analytics, the group that would later merge with Deakin Software and Technology Innovation Laboratory (DSTIL) to become Deakin Applied Artificial Intelligence Initiative. Her paper, DeepCare: A Deep Dynamic Memory Model for Predictive Medicine, was presented at the Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD) in Auckland in 2016. This year, that same paper received PAKDD’s Most Influential Paper Award. This “test of time” honour given to work whose significance has only grown over the intervening decade. It has now been cited more than 1,000 times.
“Do not always play it safe,” he says. “When the payoff comes, it can come big.” – Prof. Truyen Tran
Seeing a patient as a story, not a snapshot
What made DeepCare unusual was a deceptively simple shift in perspective. Earlier models tended to treat a patient as a collection of disconnected data points. DeepCare instead treated a person’s medical history as a living timeline, a sequence of diagnoses, treatments and hospital visits that shape one another over time.
By learning how illnesses progress, how interventions alter those trajectories, and how past events influence future risk, the model could forecast outcomes such as hospital readmission or disease relapse with a nuance earlier statistical methods couldn’t match. Crucially, it captured not just what happened to a patient but when, learning long-term patterns from the irregular, incomplete and messy data that real clinical records always are.
The concept of a health trajectory was already familiar to clinicians. What was new, Prof. Tran explains, was an AI capable of modelling it.
“There wasn’t any AI capable of doing so,” he says, nothing that could learn from ten years of a patient’s history and project that pattern far into the future.
A first paper, and a hard one
DeepCare was Dr. Pham’s very first research project, and she was, by her own account, “still very inexperienced and struggled a lot with research work.” Two problems loomed large: how to wrangle a complex, real-world medical dataset, and how to design a model that genuinely captured the structure of that data.
– Image from the paper: Figure. 2. A long-short term memory (LSTM) unit that reads input xt and previous output state ht−1 and produces current output state ht. A unit has a memory cell ct , an input gate it, an output gate ot and a forget gate ft.
She had help from a supervisory team that included Prof. Tran, then PRaDA Director and now Deakin Applied AI Co-Director, Deakin Distinguished Prof. Svetha Venkatesh, and Prof. Dinh Phung who heads Monash University’s Machine Learning Group, provided direction. But the persistence was hers. The tools for building AI a decade ago were, as Prof. Tran puts it, so poor that “most people would give up.” Pham was one of the few who didn’t.
He remembers her as one of the most skilful programmers in the group, an extraordinarily hard worker – she produced around eight papers during her PhD – and, unusually for someone of her ability, deeply coachable. “She was a rare case of being smart and skilful and, at the same time, open to feedback,” he says.
For Dr. Pham, the moments that stay with her are the breakthroughs that came after long stretches of being stuck: “when we actually found a new idea that led to positive results after a long time of being blocked.”
Why it still matters
DeepCare helped establish a new paradigm in healthcare AI, one where models understand patients through the arc of their health journeys rather than as isolated readings. That shift now underpins many systems aiming to deliver more personalised, predictive and proactive care, and the citations have come from both the AI community and clinical researchers.
The line of work also moved beyond proof of concept. Prof. Tran’s team later partnered with Victoria’s Department of Health to sort people living with diabetes into low-, medium- and high-risk groups, helping target care packages where they were needed most.
Prof. Tran is candid that the field hasn’t yet solved the hardest problem Dr. Pham’s project set out to tackle – predicting a patient’s entire future. Progress has been faster on diagnosis than on prediction, and the real barriers now are rarely technical. Health systems are conservative for good reason: clinicians must remain the ultimate decision-makers, and any AI has to be trustworthy, consistent and unobtrusive. Above all, it needs good, well-governed data, without compromising patient privacy. As Prof. Tran notes, a predictive model doesn’t need to know your name or address; it needs to know what illnesses and interventions shaped your health.
For Dr. Pham, now a senior software engineer at Google in Australia, the project’s influence has outlasted the PhD itself. The skills she built – handling complex data, understanding models deeply enough to turn an idea into working code – remain central to her work today, where she contributes data to train and evaluate AI models at Google Research.
Neither she nor her supervisor imagined a 1,000-citation future when the paper was accepted; an acceptance that was itself uncertain, precisely because the work was so different from everything around it. And that, for Prof. Tran, is the lesson worth passing to the next generation of researchers: ambition, and a willingness to take risks, can pay off in ways no one predicts. “Do not always play it safe,” he says. “When the payoff comes, it can come big.”
Header image: unsplash.com/Google DeepMind
Stories worth sharing
The text of this article is licensed under the Creative Commons Attribution (CC BY) 4.0 International license. We’d love for you to share it, so feel free. Images, videos, graphics and logos are not covered by the CC BY license and may not be used without permission from Deakin or the respective copyright holder. For more information on how to share or reuse this content, please contact researchcomms@deakin.edu.au.
Share