This seminar will be presented by Associate Professor Hanna Suominen, from the ANU Research School of Computer Science.
In health informatics, Machine Learning (ML) could be the difference between life and death when used to support decision-making in healthcare. In many clinical conditions, complex causalities and correlations between diseases and their biomarkers lead to lacking objectivity in their diagnosis, progression, and management. ML has shown its power in discovering biomarkers and allowing a continuous and burden-free remote monitoring and treatment management.
However, regardless of this promise of ML for healthcare, failures in the flow of information can lead to even life-and-death impact on patients’ lives, making this an exacerbated instance of a wider issues in computer-assisted decision-making. Yearly 500,000 Australians are rather harmed than helped in our hospitals. Yet, the issue of diagnosis and treatment delays, administration of wrong treatments or medications, missed or duplicated tests, and other preventable adverse events in healthcare is only getting worse.
Our research ambition is to contribute to technology transparency and trustworthiness that will provide substantial effectiveness and efficiency dividends, releasing clinicians time and increasing the flow of patient diagnostics. We work at the forefront of this field at the ANU Research School of Computer Science by leading the Big Data Program of Our Health in Our Hands (OHIOH) — the inaugural ANU Grand Challenge Program — and the Theory and Applications in Multimodal Pattern Analysis (TAMPA) team within the Data61 ML Group. Our approach applies and contributes to the state-of-the-art in ML by combining multi-modal data from multiple sources in a targeted manner to develop clinically useful biomarkers. The resulting more robust determination of disease status will give unprecedented information about disparate disease aspects.
To illustrate our progress, we have studied human voice as a Parkinsonian biomarker with ML detecting symptoms imperceptible to a neurologist. Our ML experiments have indicated that data featurisations used in prior literature do not perform well when extrapolated to the much larger data set. Owing to the natural variation in speech, the separation of patients and controls is not as simple as previously believed. We have presented significant performance improvements using additional novel features (with 88.6% certainty, derived from a Bayesian correlated t test) in separating patients and controls, with accuracy exceeding 58%.
Our vision when realised will enable technology-empowered workforce where patients receive earlier diagnoses and participate in enhanced disease management, regardless of their geographic location or social circumstances. It will also provide clinicians with detailed, valuable tools to help care for their patients more effectively. These contributions will drive systemic changes towards a healthier nation and societal prosperity.