van der Schaar Lab
Hub for Healthcare

Hub for Healthcare

The Hub for Healthcare matches medical problems to canonical formalisms and solutions via AI, machine learning, and operations research

Healthcare abounds with complex and difficult problems. Many of these are apparent to the clinician in the challenges they face every day in the diagnosis, treatment and management of patients, and are often defined descriptively in terms that are suitable for the profession. Our lab’s objective is to work closely with clinicians to understand these problems and create canonical formalisms from which AI, machine learning and operations research methods can provide rigorous solutions.

This page, known as the Hub for Healthcare, contains a classification of some medical problems and associated examples, and then provides formalisms and methods by which they can be solved. Many medical problems that are traditionally described in terms of clinical domain or disease can thus be reduced to a few common, canonical formalisms and, through these formalisms, solved. The intent here is to help move medicine from art to science.

The Hub for Healthcare is intended as a support resource for the clinical, AI, machine learning and operations research communities to help many of us who struggle to convert the narratives of medicine into strict formalisms that can be solved.

The table below consists of medical problems submitted by clinicians and organized by clinical domain and clinical problem type. For each of these problems, we show the  canonical formalism(s) through which these problems should be viewed and solutions using AI, machine learning and operations research that use these formalisms.

In this short video, Mihaela van der Schaar explains what we mean by “formalisms,” and why they’re an essential part of the process of solving healthcare problems through the use of AI, machine learning, and operations research.

Using the Hub for Healthcare

The Hub for Healthcare maps problems or questions raised by clinicians to the formalisms that can be used to understand and define these problems, and then to machine learning solutions. These are color-coded, allowing users to see the parts of questions that correspond to specific formalisms.

The table below is fully searchable, filterable, and sortable: you can either browse it, or—if you already know what you’re looking for—use specific keywords and phrases to find problems and discussions related to diseases, medical scenarios, or even by machine learning sub-fields. To search for a particular phrase, use the “Search” box directly above the table.

In this short video, Nick Maxfield explains how the Hub for Healthcare ties real-world medical problems and questions explicitly to formalisms, and then to solutions based on AI, machine learning, and operations research.

As of December 2020, the Hub contains problems, formalisms, and solutions primarily related to the acute care medical domain. As this project progresses (in tandem with our lab’s engagement sessions for the clinicical community), the Hub will expand to cover medical domains including cancer, cardiovascular disease, respiratory disease, organ transplantation, and many more.

Medical domain(s) Problem type(s) ML for healthcare framework Canonical formalism(s)
Sort wdt_ID Medical domain(s) Problem type(s) Raised by Description of problem ML for healthcare framework Canonical formalism(s) Solution(s)
2 34 Acute care Intervention, treatment effects Dr. Paul Elbers (NDL)

Getting the right dose to the right patient at the right time may save lives. The pharmaco-kinetic models we currently use to guide our dosing strategy consider a very limited number of covariates. How can machine learning utilize high-quality intensive care data to better inform our dosing models?

- Dr. Paul Elbers (NDL)

Causal inference/individual treatment effects

I. Bica, J. Jordon, M. van der Schaar, "Estimating the Effects of Continuous-valued Interventions using Generative Adversarial Networks," Neural Information Processing Systems (NeurIPS), 2020. [Link]

I. Bica, A. Alaa, C. Lambert, M. van der Schaar, "From real-world patient data to individualized treatment effects using machine learning: Current and future methods to address underlying challenges," Clinical Pharmacology & Therapeutics, 2020. [Link]

I. Bica, A. M. Alaa, J. Jordon, M. van der Schaar, “Estimating Counterfactual Treatment Outcomes over Time through Adversarially Balanced Representations,” International Conference on Learning Representations (ICLR), 2020. [Link]

B. Lim, A. Alaa, M. van der Schaar, “Forecasting Treatment Responses Over Time Using Recurrent Marginal Structural Networks,” Neural Information Processing Systems (NeurIPS), 2018. [Link]

2 26 Acute care, Critical care/ICU Monitoring, Treatment effects, Intervention Dr. Ari Ercole (UK)

What is the additional predictive benefit to be gained from keeping a particular patient at high risk of readmission in the ICU for another day of monitoring? If this predictive gain is small, it may be more resource-effective to institute a higher level of monitoring on the ward instead for this individual.

- Dr. Ari Ercole (UK)

Causal inference/individual treatment effects

Active sensing and value of information

I. Bica, A. M. Alaa, J. Jordon, M. van der Schaar, “Estimating Counterfactual Treatment Outcomes over Time through Adversarially Balanced Representations,” International Conference on Learning Representations (ICLR), 2020. [Link]

B. Lim, A. Alaa, M. van der Schaar, “Forecasting Treatment Responses Over Time Using Recurrent Marginal Structural Networks,” Neural Information Processing Systems (NeurIPS), 2018. [Link]

J. Yoon, W. R. Zame, M. van der Schaar, “Deep Sensing: Active Sensing using Multi-directional Recurrent Neural Networks,” ICLR, 2018[Link]

J. Yoon, J. Jordon, M. van der Schaar, “ASAC: Active Sensing using Actor-Critic Models,” Machine Learning for Healthcare Conference (MLHC), 2019. [Link]

A. M. Alaa and M. van der Schaar, “Balancing Suspense and Surprise: Timely Decision Making with Endogenous Information Acquisition,” NeurIPS, 2016. [Link] [Poster]

Methods and software available here

2 29 Acute care, Hospital medicine, Emergency medicine Triaging, Competing risks, Clinical manifestations, Comorbidities, Intervention, Capacity planning Dr. James Ray (UK)

How can machine learning provide tools that help emergency clinicians determine the severity/acuity of patients coming into the department so they can stratify emergency patients and decide what kind of emergency care is needed? Straightforward score systems used at admission time are limited in their utility, and there are a lot of risk factors (such as age, ethnicity, and obesity) that aren’t taken into account, despite being linked to vulnerability.

- Dr. James Ray (UK)

Prediction

Active sensing and value of information

Clustering

A. M. Alaa, M. van der Schaar, "AutoPrognosis: Automated Clinical Prognostic Modeling via Bayesian Optimization with Structured Kernel Learning," ICML, 2018. [Link]

A. M. Alaa, T. Bolton, E. Di Angelantonio, J. H. F. Rudd, M. van der Schaar, “Cardiovascular Disease Risk Prediction using Automated Machine Learning: A Prospective Study of 423,604 UK Biobank Participants,” PloS One, 2019. [Link]

K. Ahuja, M. van der Schaar, "Risk-Stratify: Confident Stratification Of Patients Based On Risk," submitted for publication, 2018. [Link]

C. Lee, M. van der Schaar, "Temporal Phenotyping using Deep Predicting Clustering of Disease Progression," International Conference on Machine Learning (ICML), 2020. [Link]

2 30 Acute care Diagnosis, Phenotyping, Clinical manifestations Dr. James Ray (UK)

What kind of data-driven support software can machine learning provide for differential diagnosis (including phenotyping)?

- Dr. James Ray (UK)

Clustering

K. Ahuja, M. van der Schaar, "Risk-Stratify: Confident Stratification Of Patients Based On Risk," submitted for publication, 2018. [Link]

C. Lee, M. van der Schaar, "Temporal Phenotyping using Deep Predicting Clustering of Disease Progression," International Conference on Machine Learning (ICML), 2020. [Link]

2 23 Acute care, Critical care/ICU Treatment, Risk assessment Dr. Lucas Fleuren (NLD)

Can we predict whether and when a patient on mechanical ventilation can be safely extubated? For a given patient, would the chance of successful extubation change from an additional day on the ventilator?

- Dr. Lucas Fleuren (NLD)

Causal inference/individual treatment effects (over time)

I. Bica, A. M. Alaa, J. Jordon, M. van der Schaar, “Estimating Counterfactual Treatment Outcomes over Time through Adversarially Balanced Representations,” International Conference on Learning Representations (ICLR), 2020. [Link]

B. Lim, A. Alaa, M. van der Schaar, “Forecasting Treatment Responses Over Time Using Recurrent Marginal Structural Networks,” Neural Information Processing Systems (NeurIPS), 2018. [Link]

Methods and software available here

1 24 Acute care, Hospital medicine, Perioperative medicine Personalized monitoring, Risk assessment Dr. Brent Ershoff (USA)

What is the optimal frequency of vital sign measurements for a specific patient in the hospital? For a given patient, do the observed vital signs indicate an increased risk of clinical deterioration and necessitate transferring the patient to a higher level of care?

- Dr. Brent Ershoff (USA)

Active sensing and value of information

Time series analysis and forecasting

Causal inference/individual treatment effects

J. Yoon, W. R. Zame, M. van der Schaar, “Deep Sensing: Active Sensing using Multi-directional Recurrent Neural Networks,” ICLR, 2018[Link]

J. Yoon, J. Jordon, M. van der Schaar, “ASAC: Active Sensing using Actor-Critic Models,” Machine Learning for Healthcare Conference (MLHC), 2019. [Link]

A. M. Alaa and M. van der Schaar, “Balancing Suspense and Surprise: Timely Decision Making with Endogenous Information Acquisition,” NeurIPS, 2016. [Link] [Poster]

A. M. Alaa, M. van der Schaar, “Attentive State-Space Modeling of Disease Progression,” Neural Information Processing Systems (NeurIPS), 2019. [Link]

C. Lee, J. Yoon, M. van der Schaar, “Dynamic-DeepHit: A Deep Learning Approach for Dynamic Survival Analysis with Competing Risks based on Longitudinal Data,” IEEE Transactions on Biomedical Engineering (TBME), 2019. [Link]

D. Jarrett, J. Yoon, and M. van der Schaar, “MATCH-Net: Dynamic Prediction in Survival Analysis using Convolutional Neural Networks,” NeurIPS Machine Learning for Health Workshop, 2018. – Selected as spotlight talk [Link] [Poster]

A. Bellot, M. van der Schaar, "Flexible Modelling of Longitudinal Medical Data: A Bayesian Nonparametric Approach,” ACM Transactions on Computing for Healthcare, 2020. [Link]

A.M. Alaa, M. van der Schaar, "Deep Multi-task Gaussian Processes for Survival Analysis with Competing Risks," NeurIPS, 2017. [Link]

J. Yoon, J. Jordon, M. van der Schaar, “GANITE: Estimation of Individualized Treatment Effects using Generative Adversarial Nets,” ICLR, 2018.[Link]

A. M. Alaa, M. van der Schaar, “Validating Causal Inference Models via Influence Functions,” International Conference on Machine Learning (ICML), 2019. [Link]

I. Bica, A. M. Alaa, J. Jordon, M. van der Schaar, “Estimating Counterfactual Treatment Outcomes over Time through Adversarially Balanced Representations,” International Conference on Learning Representations (ICLR), 2020. [Link]

B. Lim, A. Alaa, M. van der Schaar, “Forecasting Treatment Responses Over Time Using Recurrent Marginal Structural Networks,” Neural Information Processing Systems (NeurIPS), 2018. [Link]

Methods and software available here

2 25 Acute care, Perioperative medicine Risk assessment, Treatment effects Dr. Brent Ershoff (USA)

Given a patient's preoperative history of either chronic pain or opioid use, can we determine which subgroup of patients may benefit from either an inpatient acute care pain consult or treatment with nonopioid adjunctive medications?

- Dr. Brent Ershoff (USA)

Clustering

Causal inference/individual treatment effects

C. Lee, M. van der Schaar, "Temporal Phenotyping using Deep Predicting Clustering of Disease Progression," International Conference on Machine Learning (ICML), 2020. [Link]

C. Lee, J. Rashbass, M. van der Schaar, "Outcome-Oriented Deep Temporal Phenotyping of Disease Progression,” accepted to IEEE Transactions on Biomedical Engineering, 2020.

A.M. Alaa, M. van der Schaar, "Deep Multi-task Gaussian Processes for Survival Analysis with Competing Risks," NeurIPS, 2017. [Link]

J. Yoon, J. Jordon, M. van der Schaar, “GANITE: Estimation of Individualized Treatment Effects using Generative Adversarial Nets,” ICLR, 2018.[Link]

A. M. Alaa, M. van der Schaar, “Validating Causal Inference Models via Influence Functions,” International Conference on Machine Learning (ICML), 2019. [Link]

I. Bica, A. M. Alaa, J. Jordon, M. van der Schaar, “Estimating Counterfactual Treatment Outcomes over Time through Adversarially Balanced Representations,” International Conference on Learning Representations (ICLR), 2020. [Link]

Methods and software available here

2 33 Acute care, Critical care/ICU Early warning systems, Risk assessment, Capacity planning, Prognosis, Treatment effects, Personalized Monitoring Dr. Steve Harris (UK)

How do we identify the point at which we transition from non-invasive to invasive ventilation? On one hand, intensive ventilation is not a decision to take lightly; on the other hand, lack of invasive ventilation may in some cases lead to self-induced lung injury.

- Dr. Steve Harris (UK)

Time series analysis and forecasting

Causal inference/individual treatment effects

A. M. Alaa, M. van der Schaar, “Attentive State-Space Modeling of Disease Progression,” Neural Information Processing Systems (NeurIPS), 2019. [Link]

I. Bica, A. M. Alaa, J. Jordon, M. van der Schaar, “Estimating Counterfactual Treatment Outcomes over Time through Adversarially Balanced Representations,” International Conference on Learning Representations (ICLR), 2020. [Link]

B. Lim, A. Alaa, M. van der Schaar, “Forecasting Treatment Responses Over Time Using Recurrent Marginal Structural Networks,” Neural Information Processing Systems (NeurIPS), 2018. [Link]

Methods and software available here

2 32 Acute care, Critical care/ICU, Hospital medicine, Emergency medicine Early warning systems, Risk assessment, Resource optimization Dr. Thom Daniels (UK)

Existing early warning score systems are blunt tools: they only look at data from a single time point, exclude pathology data or demographics, and don't learn over time. How can machine learning improve early warning scores using all the data clinicians have access to?

- Dr. Thom Daniels (UK)

Time series analysis and forecasting

A. M. Alaa, M. van der Schaar, "AutoPrognosis: Automated Clinical Prognostic Modeling via Bayesian Optimization with Structured Kernel Learning," ICML, 2018. [Link]

J. Yoon, A. M. Alaa, S. Hu, M. van der Schaar, “ForecastICU: A Prognostic Decision Support System for Timely Prediction of Intensive Care Unit Admission,” ICML, 2016. [Link]

Methods and software available here

2 31 Acute care Diagnosis, Imaging, Resource optimization Dr. James Ray (UK)

Given the amount of time tests and scans can take to process, how can machine learning analyze patterns within tests or scans and interpret them more quickly?

- Dr. James Ray (UK)

Prediction

Operations research

T. Kyono, F. J. Gilbert, and M. van der Schaar, "MAMMO: A Deep Learning Solution for Facilitating Radiologist-Machine Collaboration in Breast Cancer Diagnosis," 2018. [Link]

A. M. Alaa, T. Bolton, E. Di Angelantonio, J. H. F. Rudd, M. van der Schaar, “Cardiovascular Disease Risk Prediction using Automated Machine Learning: A Prospective Study of 423,604 UK Biobank Participants,” PloS One, 2019. [Link]

C. Tekin, O. Atan and M. van der Schaar, “Discover the Expert: Context-Adaptive Expert Selection for Medical Diagnosis,” IEEE Transactions on Emerging Topics in Computing, vol. 3, no. 2, pp. 220-234, June 2015. [Link]

Note: many of the machine learning algorithms referenced in this table are available as software packages that can be found in our BitBucket repository.

Our Clinical Advisory Groups

Machine learning for medicine is an inherently interdisciplinary area, and cannot be successfully approached unilaterally from the perspective of machine learning.

Our lab has been fortunate enough to benefit from expert guidance and collaborative support from an international network of doctors who are leaders in their respective domains of specialization. Some of our collaborators have kindly agreed to serve as members of Clinical Advisory Groups for medical subdomains covered in the Hub for Healthcare.

Each Clinical Advisory Group member helps us to define the scope and boundaries of their subdomain (as well as medical problem types within that domain), and advises us regarding the problems encountered by clinicians working in their domain.

Acute Care Clinical Advisory Group

Dr. Maxime Cannesson

Chair, UCLA Department of Anesthesiology & Perioperative Medicine
University of California, Los Angeles, USA

Hub for Healthcare - Dr. Maxime Cannesson
Twitter
Google Scholar
Email

Dr. Maxime Cannesson is the Ronald L. Katz, MD, Chair of the UCLA Department of Anesthesiology and Perioperative Medicine. His department engages in research, teaching, and patient care activities across Los Angeles.

Sites include Ronald Reagan UCLA Medical Center in Westwood, UCLA Medical Center in Santa Monica, the Resnick Neuropsychiatric Hospital, Mattel Children’s Hospital, and the UCLA Medical Group with its extensive system of primary-care and specialty-care offices throughout the region. Consistently ranked one of the top ten hospitals in the nation – and the best in Los Angeles, UCLA is at the leading edge of medical care and biomedical research.

Dr. Cannesson’s department currently has four main research themes, each of which is supported by a strong foundation of basic and clinical research infrastructure: cardiovascular research; neuroscience, mechanism of anesthesia and pain; organ-protection research; and biocomputing and health informatics.

Dr. Paul Elbers

Intensivist at Amsterdam UMC, the Netherlands

Hub for Healthcare - Dr. Paul Elbers
Twitter
LinkedIn
Google Scholar
Email

Dr. Paul WG Elbers is an intensivist at VU medical center, Amsterdam, The Netherlands.

Special interests include quantitative acid base analysis, fluid therapy, ultrasound, pharmacokinetics, medical information technology, and human microcirculation.

The latter was the topic of his PhD thesis entitled “Focus on Flow, Imaging the Human Microcirculation in Periopera-tive and Intensive Care Medicine”.

Together with Prof. John A Kellum he edited the second edition of “Stewart’s Textbook of Acid-Base”. Together with Rainer Gatz, MD, he is the editor of www.acidbase.org, an online community for quantitative acid base analysis including clinical decision support tools.

Paul is one of the organizers of the Amsterdam Medical Data Science Group, which plays a leading role in trying to solve the ethical and moral dilemma of data sharing by providing guidance for European Intensive Care Units to share their data in a responsible way. This work is being done with the European Society of Intensive Care Medicine.

Dr. Ari Ercole

Consultant in Neurointensive Care,
Cambridge University Hospitals NHS Foundation Trust, UK

Hub for Healthcare - Dr. Ari Ercole
Twitter
LinkedIn
GitHub
Google Scholar
Email

Dr. Ari Ercole became an anaesthetist after completing a PhD in physics at the University of Cambridge. He divides his time between intensive care, anaesthesia and research.

His particular focus on data-driven research: Developing novel analytical techniques including machine learning and feature discovery for intensive-, acute- and perioperative care.

Dr. Ercole has an extensive background in physical and statistical modelling as well as data science and computing. His research involves development of novel medical technology (in particular data analysis techniques but also sensor development and physiological measurement) and its application to the treatment of critically ill patients in particular. He is interested in physiological signal complexity (fractal dynamics, signal entropy) as an emergent behaviour of networks of non-linear dynamical systems or techniques from machine learning and time-series statistics to neurophysiological and clinical data to detect and characterise novel multidimensional interrelationships.

Dr. Brent Ershoff

Assistant Professor-In-Residence, Department of Anesthesiology
David Geffen School of Medicine, University of California, Los Angeles, USA

Hub for Healthcare - Dr. Brent Ershoff
LinkedIn
Google Scholar
Email

Following a BS and MS in Biological Sciences, both at Stanford University, Dr. Ershoff received his MD from University of California, San Francisco. This was followed by an anasthesia residency at UCLA Health.

Dr. Ershoff completed a liver transplant anesthesiology fellowship at UCLA Health in 2015, and is now Assistant Professor-In-Residence at UCLA’s David Geffen School of Medicine.

Cardiovascular Disease Clinical Advisory Group

Cancer Clinical Advisory Group