[edit]
Healthcare AI: Key to better patient-practitioner-system partnerships
[edit]
Room: Yaiza
Machine learning in healthcare has focused on surfacing useful structure from high-volume, sometimes high-velocity, data from technologies such as imaging and 'omics. As this structure is translated into actionable information the biomedical ambitions for machine learning are growing. The crossover between diagnosis and therapy, in cancer radiotherapy for example, is linking observation and intervention processes for more precise care. In closely coupled observation-intervention scenarios, such as brain stimulation linked to exoskeleton-guided stroke rehabilitation, AI may open new, multi-component therapies. Single diseases, disciplines, mechanisms and healthcare technologies attract narrowly-defined data, learning and inference. Yet, the major burden of ill-health and healthcare uncertainty lies between these foci, in the noisy, real-world complexity of health and care. Some AI investments have started to tackle inference problems such as hastening interventions to save kidney function via inferences from sparse clinical data on patients with a wide variety of conditions. Indeed, conditions such as high blood pressure and diabetes interact, worsening decline in kidney function. These two conditions and drug x may also interact, but in ways not studied in trials of drug x where patients with multiple conditions were excluded. In the rhythms of everyday life, key determinants of drug outcomes rest with the patient, for example diet and physical activity alongside diabetes medication. Arguably the greatest challenge for AI in healthcare is to surface the dynamics of health systems, hastening and targeting interventions. This workshop will assemble a variety of perspectives in accelerating, assisting and augmenting patients, practitioners and health system managers with machine learning. We will challenge ourselves to ask: Are we asking the right questions for maximum societal impact from health-related machine learning? Are we borrowing strength sufficiently across silos of inference? How does this community organise at global scale to share well-formed problems to advance methodology and avoid AI hype.09:30-11:00 Challenge 1: Are we asking the right questions for maximum societal impact from health-related machine learning? Fitting more complex functions to a previously published objective is not a good use of machine learning effort; problems need better selection, framing and evaluation metrics.
- Chair:
- Patient-centred research in a complex health data world: a new role for ML.
- Learning from patient-contributed social and mobile data: for the patients, the researchers, the clinicians, or for all?
- Going the last mile: why deep learning is not enough.
11:00-11:30 Coffee Break
11:30-13:00 Challenge 2: Are we borrowing strength sufficiently across silos of inference to be patient-centred and population-serving? Studies that focus on prediction using offline datasets without collaborative feedback are limited; problems need addressing within systems at reasonable resolution (time, place, person).
- Chair:
- Smartphone-based digital phenotyping.
- Adapting AI approaches to rare diseases.
- Doctors and Interpreters: Interpretability and Visualization in the Medical Domain.
13:00-14:00 Lunch
18:00-18:45 Challenge 3: How does this community organise at global scale to share well-formed problems to advance health data science and avoid AI hype? Technology-led collaborations attract poorly framed problems; our field needs to embrace a more interdisciplinary approach and pursue more careful and meaningful collaborations, in concert.
- Chair:
- Democratisation of data science.
- Collaboration as an innovation to advance healthcare AI.
18:45-19:30 Next steps and closing remarks
- Chair:
- AMIE panel: Riccardo Bellazzi, Anita Burgun, John Holmes, Suchi Saria and Allan Tucker