Machine Learning Opportunities for Healthcare

CDS’ Rajesh Ranganath, with researchers from MIT, Harvard, and Johns Hopkins, explores the future for healthcare applications of machine learning

While machine learning applications face challenges in every domain, healthcare presents the most significant hurdles — but patients can benefit in critical ways from machine learning models in clinical scenarios. Researchers recently offered a new outline of the particular obstacles to and benefits of applying machine learning to healthcare. Authors of the outline include Rajesh Ranganath, Assistant Professor of Computer Science and Data Science, Marzyeh Ghassemi, MIT, Verily, and the University of Toronto, Tristan Naumann, MIT, Peter Schulam, Johns Hopkins University, and Andrew L. Beam, Harvard Medical School.

The researchers focused on applications of machine learning for electronic health records (EHRs), which contain daily operational, clinical, and financial data. The researchers established a new taxonomy for EHRs for the purpose of machine learning: high-frequency monitors at a patient’s bedside, biomarkers from vitals and lab tests, and notes that document interactions between a patient and a healthcare team.

Many potential healthcare applications for machine learning involve using data from EHRs to determine the outcomes of various treatment methods. But building models to address such causal problems is most difficult when datasets are comprised of observational data — the type of data contained in EHRs — because interventions in observational data may not be randomly assigned. Consequently, it is imperative for machine learning methods to adjust for the presence of confounders to provide estimates of outcomes under various treatments.

Labels also need to be continually updated since medical definitions evolve due to medical advances. The researchers emphasized the importance of continuous training for machine learning models in healthcare — even when deployed, the models need to continue to train on new, changing data. An important concern in deploying models is that models trained on biased data would amplify existing bias in medical practice.

Despite the challenges, there are clear opportunities for machine learning intervention in healthcare. Ranganath and collaborators point specifically to:

  • Automation tasks including medical image evaluation and routine processes such as triage prioritization;
  • Clinical support and augmentation including standardizing medication prescription and dosage and integrating fragmented records to identify at risk patients like domestic abuse victims from repeated hospital visits;
  • Early detection of medical events like hypertension or senior falls with wearable medical devices;
  • Development of precision medicines for individualized treatment, which is crucial for patients with syndromes — medical conditions defined by a collection of symptoms whose causes are unknown.

Ranganath and collaborators also point out that most commonly used ICU treatments are not backed by evidence from randomized control trials (RCTs), but machine learning techniques on observational data can support RCT results to help identify individual treatment pathways.

The researchers stress machine learning models for healthcare have more utility when they justify particular predictions for medical practitioners. Fundamentally, these models need to interact and collaborate with humans in interpretable ways — more so than in other domains — to benefit patients and caregivers.

By Paul Oliver