Causal Machine Learning for Healthcare and Precision Medicine

In plain words

Most machine learning in medicine learns which patterns go with which outcomes. That is useful for spotting a disease, but it does not say what would happen to a particular patient if they got drug A instead of drug B, because a patient record only shows the choice that was actually made. This paper argues that doctors' decision support needs causal reasoning, which models how the data came to be, and it reviews the research directions that could bring that reasoning into machine learning.

The running example is Alzheimer's disease. A brain scan can look older because the patient is older, not because of the disease, so a classifier that ignores age does worse on older healthy people. Modern health data is also mixed (images, text, genetics), changes over time, and comes from different hospitals, so models that learn shortcuts fail when they move.

The authors sort the field into three directions: learning causal variables from raw high-dimensional data, discovering which variables cause which, and reasoning about what an intervention would change. They are candid that these problems are open. The paper is a review with one illustrative table, not a method with code.

TL;DR

Review and perspective, not a new method. It argues that precision medicine needs individual treatment effects, which observational data cannot show directly, so causal assumptions are required §1 §3.3. Its one table is a replication of earlier work and shows that counterfactual augmentation raises precision for the oldest group but lowers recall for it tab. 1. No code and no method hyperparameters are given; use it as a map of the field.

Problem

• Clinical decision support models mostly learn associations and cannot tell causes from spurious correlations §1.
• Choosing a treatment for one patient needs the difference between the outcome with and without treatment. Only one of the two is ever observed, which the paper calls the "fundamental problem of causal inference" §2.2 §3.3.
• In observational data, treatment choice and outcome both depend on the patient's condition (confounding). A randomised controlled trial removes this, but is not always feasible §2.3 §3.3.
• An average treatment effect can hide large differences between patients, so the aim is an individual or conditional effect §3.3.
• Health data is multimodal, high-dimensional, unstructured and temporal §1 §4.1 §4.2.
• Models trained at one site can fail at another when the environment shifts (scanner, cohort, hospital). A spurious link such as scanner type can stand in for disease status §4.3.
• Causal conclusions need assumptions. None can be learned from observational data alone §2.

Method

This is a review. It defines the framework used throughout and lists open directions. Nothing here is a training procedure the paper evaluates.

Framework. The paper splits causal machine learning into three directions §1 fig. 1:

1. Causal representation learning: from high-dimensional data, learn low-dimensional causal variables and the relations between them §5.1.
2. Causal discovery: from a set of variables, learn their causal relations §5.2.
3. Causal reasoning: given variables and their relations, analyse how the system reacts to interventions and counterfactual queries §5.3.

Structural causal models. A model is a set of structural assignments X_k := f_k(\mathbf{PA}_k, N_k), one per variable, where \mathbf{PA}_k are the direct causes and N_k is noise §2.1 eq. 1. The noise terms are independent, and the joint distribution factorises into one conditional per variable §2.1 eq. 2:

P(X_1,\dots,X_K) = \prod_{k=1}^{K} P(X_k \mid \mathbf{PA}_k) eq. 2

The paper lists three levels of query: association, intervention do(\cdot) and counterfactual §2.1. Pearl's causal hierarchy names these levels §2.1.

Potential outcomes. Each unit has an outcome under control, Y^{(0)}, and under treatment, Y^{(1)}. The effect is their difference §2.2. The average effect is E[Y^{(1)} - Y^{(0)}], usually estimated as E[Y \mid T=1] - E[Y \mid T=0], and the individual effect is \tau_i = Y_i(1) - Y_i(0). The paper uses the term CATE, a conditional average, for the estimable version of the individual effect §3.3.

Meta-learners. The paper describes this family as the most common route to CATE from observational data §3.3. The paper gives no algorithm box for it alg. 1 not stated. The pseudocode below restates that description.

# Meta-learner estimate of CATE  [§3.3]
# Input:  features (N, d); treatments (N,) in {0, 1}; outcomes (N,)
# Output: a function tau_hat(x) that returns one value per input

def estimate_cate(features, treatments, outcomes):
    # Split units by arm; the paper does not name a learner  [§3.3]
    control = treatments == 0
    treated = treatments == 1
    # Predict the outcome of control units from their features  [§3.3]
    mu_control = fit_regressor(features[control], outcomes[control])
    # Predict the outcome of treated units from their features  [§3.3]
    mu_treated = fit_regressor(features[treated], outcomes[treated])

    def tau_hat(x):
        # The difference of the two predictions estimates CATE  [§3.3]
        return mu_treated(x) - mu_control(x)

    return tau_hat

Identification assumptions. Estimating a treatment effect from observational data needs conditional exchangeability (no unmeasured confounders), positivity (every patient could receive each treatment) and consistency (the treatment is defined unambiguously) §3.3. The paper says that measuring more variables can make exchangeability more plausible §4.1.

Counterfactual augmentation. The paper's illustration fixes a patient's Alzheimer's status and changes age, and the reverse, using a conditional generative model §3.1 fig. 2. Those synthetic images are then added to the classifier's training data, so that the classifier separates the effect of ageing from that of disease §3.2. The paper does not state how many images are generated or what label they carry.

Causal representation learning. Model the representation domain as a structural causal model, building on disentangled representation learning but with stronger inductive bias §5.1.

Causal discovery. Most methods rely on conditional independence tests, combinatorial search over DAGs, or assumptions about function class and noise (linear, additive, Gaussian) §5.2.

Causal reasoning. Counterfactual image features may need invertible mechanisms such as normalising flows, or variational inference, which have their own costs §5.3.

Learning settings. Supervised learning assumes one environment for training and test. Domain adaptation has unlabelled target data in training. Domain generalisation targets unseen environments. The paper also lists semi-supervised, continual, fairness and imbalanced settings through the same causal lens §4.3.

Hyperparameters

Results

Author-reported, and replicated from earlier work. Setup: ADNI scans, a convolutional classifier for Alzheimer's diagnosis, precision and recall on the test set, stratified by age §3.2 tab. 1. Baseline: naively trained classifier. Comparison: classifier trained with counterfactually augmented data. The paper states this is a replication of Xia et al. (2022) and not an extension of it §3.2 fn.

The paper reads these as the naive classifier doing worst on healthy older patients (80-90), and augmentation improving precision for that group tab. 1 §3.2. The paper gives no sample sizes, confidence intervals or significance tests [not stated]. The paper reports no other quantitative results. The synthetic-image example is qualitative §3.1 fig. 2; the paper points to Xia et al. for quantitative confirmation §3.1 fn.

Paper vs code

No official code. The abstract and the LaTeX source contain no repository link and no project page. A GitHub search for "causal machine learning healthcare" found only unrelated student and portfolio projects. The last author's lab organisation (vios-s) was searched for "causal" and found no repository for this review. The vios-s result for "causal" is a different paper's code (DiffAN, causal discovery), so it was not used.

Pitfalls

• Nothing to implement, except one illustration. The card's method sections restate definitions and one meta-learner description. An agent should not expect a training recipe here.
• The only table is not new. It is replicated from Xia et al. (2022), and the paper says it is not an extension §3.2 fn. The paper gives no sample sizes or repeat runs, so small differences (for example 87.7 vs 88.3 precision) cannot be judged as real from this paper alone.
• Mixed signal in the table. Augmentation raises precision in every age group, but recall falls for 60-70 (92.5 to 91.5) and for 80-90 (97.1 to 95.7) and rises only for 70-80 tab. 1. The caption and text highlight the precision gain for the worst group only.
• Synthetic-image figure is an illustration. The generator's training and settings are in Xia et al., not here §3.1 fn. The paper's claim that ageing and Alzheimer's change brain volume similarly rests on that work §3.1 fn.
• Causal graph is a modelling choice. The paper treats the MR image as an effect of disease and age, and age as a confounder, with the argument that age cannot be caused by brain changes §3.1. This is an assumption the paper argues for, not one it tests.
• Terms used as synonyms. The paper uses ITE and CATE for the same idea §3.3, and "conditional exchangeability" and "causal sufficiency" as one assumption §4.1. Keep the names apart in code and state which quantity is estimated.
• Individual effects are never observed. The meta-learner estimates a conditional average from groups, not a per-patient effect. The paper's point that most patients may not respond like the average §3.3 means the estimate can still be wrong for a given patient.
• Temporal causality is largely open. The paper calls machine-learning temporal causal inference for disease progression largely unexplored and cites a time-series model that defines cause by predictive power of the past §4.2. No method or evaluation is given.
• Unstated design details. Architecture, optimiser, learning rate, epochs, training data size and split are absent for the classifier and the generator §3.2 §3.1.

Lineage

• Builds on: Pearl's structural causal models §2.1 and Rubin's potential outcomes §2.2, neither of which has a card in this wiki.
• Replicates: Xia et al. (2022) on counterfactual augmentation for age-confounded Alzheimer's classification §3.2 fn. Their arXiv id is not stated in this paper; no card exists here.
• Compared against: not applicable (review).
• Superseded by: not stated.