[Paper Review] A Review of Challenges and Opportunities in Machine Learning for Health

Marzyeh Ghassemi, Tristan Naumann+4 more|PubMed|Jun 1, 2018

Machine Learning in HealthcareComputer Science123 references253 citations

TL;DR

This paper surveys the unique challenges of applying machine learning to health data (especially EHRs) and outlines opportunities, emphasizing causality, missing data, outcome definitions, and non-stationarity, with a call for clinician–ML collaboration.

ABSTRACT

Modern electronic health records (EHRs) provide data to answer clinically meaningful questions. The growing data in EHRs makes healthcare ripe for the use of machine learning. However, learning in a clinical setting presents unique challenges that complicate the use of common machine learning methodologies. For example, diseases in EHRs are poorly labeled, conditions can encompass multiple underlying endotypes, and healthy individuals are underrepresented. This article serves as a primer to illuminate these challenges and highlights opportunities for members of the machine learning community to contribute to healthcare.

Motivation & Objective

• Highlight the unique technical challenges of ML in healthcare (causality, missingness, outcomes) and how they affect modeling choices.
• Present a hierarchy of healthcare opportunities where ML can automate tasks, support clinicians, and expand clinical capacities.
• Encourage collaboration between ML researchers and clinicians to develop clinically useful and operationally feasible models.
• Discuss research directions in handling data shifts, interpretability, and representation learning in healthcare

Proposed method

• Discuss causality as a core requirement for answering intervention-based questions in healthcare.
• Explain missing data mechanisms (MCAR, MAR, MNAR) and their implications for model design and evaluation.
• Outline reliable outcome construction and the risk of label leakage in EHR-based learning.
• Propose a hierarchy of healthcare opportunities and map ML methods to automate tasks, support decision-making, and expand capacities.
• Advocate for robust, interpretable, and collaborative ML systems and representational learning for multi-source healthcare data.

Experimental results

Research questions

• RQ1 What are the core challenges unique to applying ML to healthcare data and how do they affect model validity and utility?
• RQ2 How should outcomes be defined and labeled when designing ML tasks with heterogeneous EHR data?
• RQ3 How can ML models account for missing data and biases inherent in healthcare datasets?
• RQ4 What are the opportunities and requirements for deploying ML in clinical settings, including interpretability and collaboration with clinicians?
• RQ5 What research directions address data non-stationarity and representation in healthcare ML?

Key findings

• Causality is essential for answering intervention-based questions and poses challenges when using observational healthcare data.
• Missing data mechanisms (MCAR, MAR, MNAR) must be modeled and acknowledged to avoid biased predictions and misinterpretation.
• Outcomes in healthcare require careful definition, context awareness, and avoidance of label leakage to ensure meaningful predictions.
• There is a broad, three-tier opportunity structure: automating tasks, supporting clinical decisions, and expanding clinical capacities, each with distinct evaluation needs.
• Model interpretability and justifiability, along with robust representations for multi-source data, are critical for clinical adoption and trust.
• Clinical collaboration is essential to identify high-impact problems and ensure operational feasibility of ML solutions.