Source-linked AI summary
Causal Machine Learning for Healthcare and Precision Medicine
Pedro Sanchez, Jeremy P. Voisey, Tian Xia, Hannah I. Watson, Alison Q. ONeil, Sotirios A. Tsaftaris
TL;DR
Healthcare CDS systems must move beyond associations to answer individualised intervention questions for precision medicine. This paper reviews how causal machine learning combines causal inference and machine learning, then identifies unresolved data, generalisation, and temporal challenges while outlining three research directions.
Problem
Precision-medicine CDS must predict individual responses to interventions, whereas conventional machine learning primarily learns associations that may not provide actionable treatment comparisons.
Method
The paper reviews how causal inference can be incorporated into clinical decision support using recent machine-learning advances, illustrated with Alzheimer’s disease examples.
Results
The paper identifies causal representation learning, causal discovery, and causal reasoning as research directions for addressing healthcare CDS challenges.
Takeaways & Limitations
A causal perspective supports reasoning about intervention effects at population and individual levels, closing the gap toward precision medicine.
Takeaways & Limitations
Purely observational data provide limited evidence for determining causal relationships, which often requires carefully designed experiments.
Abstract
from arXiv · showhide
Causal machine learning (CML) has experienced increasing popularity in healthcare. Beyond the inherent capabilities of adding domain knowledge into learning systems, CML provides a complete toolset for investigating how a system would react to an intervention (e.g.\ outcome given a treatment). Quantifying effects of interventions allows actionable decisions to be made whilst maintaining robustness in the presence of confounders. Here, we explore how causal inference can be incorporated into different aspects of clinical decision support (CDS) systems by using recent advances in machine learning. Throughout this paper, we use Alzheimer's disease (AD) to create examples for illustrating how CML can be advantageous in clinical scenarios. Furthermore, we discuss important challenges present in healthcare applications such as processing high-dimensional and unstructured data, generalisation to out-of-distribution samples, and temporal relationships, that despite the great effort from the research community remain to be solved. Finally, we review lines of research within causal representation learning, causal discovery and causal reasoning which offer the potential towards addressing the aforementioned challenges.
1 Introduction
Healthcare clinical decision support increasingly uses machine learning, but precision medicine requires causal predictions about intervention outcomes rather than associations alone. This paper reviews how causal machine learning can support personalised decisions and address multimodal, temporal, and deployment-generalisation challenges.
- Clinical decision support: Machine-learning CDS systems predict tasks such as detection, classification, and segmentation from multimodal electronic health record data.Inputs include medical images, clinical free-text notes, blood tests, and genetic data.
- Clinical decision support: Precision-medicine CDS must estimate how an individual patient would respond to alternative interventions, such as treatment effects on Alzheimer’s cognitive outcomes.Actionable information compares outcomes under treated and untreated scenarios for the same patient.
- Healthcare challenges: A major healthcare challenge is acquiring causal information from multimodal, high-dimensional, unstructured, and temporally progressive data that generalises across deployment environments.The stated deployment environments include different hospitals, cities, and countries.
- Causal machine learning: Combining causal inference with machine learning can process complex data while incorporating expert knowledge to improve multimodal information merging, generalisation, and explainability.Machine learning contributes nonlinear modelling, while causal inference adds system-level understanding.
- Causal machine learning: The paper categorises CML research into causal representation learning, causal discovery, and causal reasoning for addressing healthcare challenges.These directions respectively extract informative causal variables, learn causal relationships, and analyse responses to interventions.
- Paper scope: The paper reviews causal concepts, healthcare applications, clinical decision-support challenges, and research directions for improving personalised decision making.The review uses Alzheimer’s disease examples to illustrate these topics.
2 What is causality?
Causality concerns directional dependence and enables reasoning about interventions, extending machine learning beyond associative prediction. The section introduces structural causal models, potential outcomes, and experimental requirements for identifying causal effects.
- Causal concepts: Causality is directional: an effect relies on its cause, but the cause does not thereby rely on the effect.This definition supports analysing how systems respond to interventions.
- Causal concepts: Clinical causal questions ask how treatments affect disease progression, counterfactual outcomes, or the best drug for an individual patient.The Alzheimer’s example compares drugs by expected cognitive decline over five years.
- Causal concepts: Causal inference requires assumptions that may be supplied by study design or contextual knowledge but cannot be learned solely from observational data.This limits what standard machine-learning approaches can identify from observations alone.
- Structural causal models: Structural causal models represent variables through structural assignments, parent variables, and exogenous noise, with a corresponding directed acyclic graph.Parent variables are direct causes, while noise models uncertainty and omitted variables.
- Structural causal models: The structural causal framework assumes autonomous causal mechanisms and mutually independent exogenous variables, yielding a factorisation into conditional mechanisms.The paper describes this as the independent causal mechanisms principle.
- Causal queries: Causal models support associative, interventional, and counterfactual queries, with do(A) fixing A so only its descendants change.This distinguishes interventional distributions from observational distributions.
- Potential outcomes: The potential-outcomes framework represents treated and untreated responses for a patient, defining treatment effect as their difference.Both potential outcomes cannot be observed simultaneously, creating the fundamental problem of causal inference.
- Determining cause and effect: Causal relationships often require carefully designed experiments because purely observational data provide limited evidence about causation.Randomised controlled trials remove confounding by randomising treatment assignment.
3 Why should we consider a causal framework in healthcare?
Causal machine learning supports healthcare decisions by modeling interventions, confounding, and heterogeneous treatment effects rather than relying only on observed associations. The Alzheimer’s disease examples show how causal models can expose bias, guide data generation and augmentation, and motivate more robust precision-medicine systems.
- Treatment Effect and Precision Medicine: Patient-level treatment decisions require estimating factual and counterfactual outcomes, making treatment-effect reasoning fundamentally different from standard supervised prediction.In observational data, the counterfactual outcome is never observed, so identifying an individual’s best treatment requires causal reasoning.
- Alzheimer’s Disease practical example: Causal graphs represent domain relationships that identify confounders and guide which variables to control when estimating intervention effects.In the Alzheimer’s example, chronological age is modeled as a confounder of brain morphology and Alzheimer’s diagnosis.
- Modeling the Data Generation Process: Conditional generative models can synthesize brain MR images while holding one variable fixed and intervening on age or Alzheimer’s disease status.The example uses a 64-year-old cognitively normal baseline and generates images under altered ages or disease statuses.
- Modeling the Data Generation Process: Intervening independently on age and Alzheimer’s disease produces similar brain-volume changes, illustrating how confounding can affect image-based diagnosis.A classifier may therefore perform worse for older cognitively normal adults because ageing produces Alzheimer’s-like brain patterns.
- Modeling the Data Generation Process: Causal knowledge can improve generalisation by linking observed shifts, such as scanner differences, to data-generation mechanisms and motivating targeted augmentation.Causal diagrams can also represent population, acquisition, and annotation shifts, as well as data scarcity.
- Treatment Effect and Precision Medicine: Individualised treatment effects can differ substantially from average treatment effects, motivating personalised recommendations rather than population-average decisions.The paper defines the individualised effect for patient i as τ_i = Y_i(1) − Y_i(0).
4 Causal machine learning for complex data
Healthcare CML must handle multimodal, high-dimensional, unstructured, and temporal data while remaining robust to distribution shifts. A causal perspective frames shortcut learning and supports invariant prediction across environments.
- Complex healthcare data: Healthcare data combine imaging, genomics, clinical assessments, demographics, and free-text reports, making multimodal integration and meaningful feature extraction difficult.Different modalities must be transformed into a shared space while disentangling common from modality-specific information.
- Temporal information: Longitudinal treatment decisions require models to represent patient history, treatment timing, and trajectories under alternative future interventions.Disease progression and treatment effects can differ across patients with different covariates.
- Out-of-distribution generalisation: Out-of-distribution generalisation concerns test distributions that differ from training distributions, often because environments such as hospitals change.The paper presents a causal perspective intended to unify such machine-learning settings.
- Out-of-distribution generalisation: Causal relationships are treated as stable across environments, so causal learning aims for predictions invariant to distribution shifts.This perspective motivates using expert-provided structure to improve robustness when collecting multiple independent datasets is costly.
- Causal perspective on prediction: In anti-causal prediction, models infer a generating factor from high-dimensional observations, but environment-dependent associations can alter P(Y|X).The causal mechanism P(X|Y) is distinguished from the prediction direction, whose distribution may change across environments.
- Causal perspective on prediction: Shortcut learning occurs when a predictor uses a spurious feature correlated with the target, such as MRI scanner type correlated with Alzheimer’s diagnosis across hospitals.The model may learn scanner type rather than the required relationship between brain images and diagnosis.
5 Research Directions in Causal Machine Learning
The paper reviews causal representation learning, causal discovery, and causal reasoning as research directions for applying CML to complex healthcare data. Each direction addresses a different stage from extracting variables and relations to evaluating interventions, while retaining substantial open challenges.
- 5.1 Causal Representations: Causal representation learning extracts informative low-dimensional variables from high-dimensional data and models their causal relationships.It extends representation learning by treating the representation domain as a structural causal model.
- 5.2 Causal Discovery: Causal discovery learns causal directions among known variables, especially when experiments are expensive, unethical, or impossible and data are observational.Its methods commonly use conditional independence tests, DAG searches, or assumptions about functions and noise.
- 5.2 Causal Discovery: Causal discovery remains limited by unobserved confounding, functional-form misspecification, and difficulty modeling temporal causal relationships.These challenges affect discovery of causal effects from observational data.
- 5.3 Causal Reasoning: Causal reasoning uses causal models to quantify intervention effects and answer counterfactual questions about alternative treatments.In healthcare, this includes asking what would have happened if a patient had received another treatment.
- 5.3 Causal Reasoning: Causal models can improve transparency by making assumptions explicit, and architectures that mirror assumed causal graphs can support interpretability.These benefits are especially relevant where communicating assumptions and explanations matters.
- 5.3 Causal Reasoning: Causal reasoning with ML remains difficult for interventions on complex representations and functions, including images, text, age, sex, and genetic data.Counterfactual image prediction may require invertible mechanisms or variational inference, each introducing additional complexity.
6 Conclusion
The paper argues that healthcare systems should model data-generation processes rather than focus only on predictive accuracy. It identifies multimodal, temporal, high-dimensional, unstructured, and cross-environment decision-making as challenges for which three CML directions offer research opportunities.
- 6 Conclusion: Causal ML aims to support reasoning about intervention effects at both population and individual levels, advancing precision medicine beyond predictive performance alone.The conclusion frames intervention reasoning as a central healthcare objective.
- 6 Conclusion: Key healthcare challenges include using multimodal, high-dimensional, and unstructured data while generalising across environments and incorporating temporal information.The paper proposes causal representation learning, causal discovery, and causal reasoning as opportunities to address them.