When designing personalized therapies, machine learning models and clinical trials share a common goal: identifying which patients will benefit most from a specific intervention. To do this, algorithms must parse complex biological data—such as gene expression profiles, medical images, or blood panels—to find features that correlate with positive patient outcomes.
However, these systems frequently fall into a subtle but pervasive trap known as prognostic-predictive confounding. Because a patient's final health outcome is a combination of their baseline disease trajectory and their specific response to a drug, standard predictive models often mistake general indicators of health or tumor aggressiveness for genuine evidence of a drug's efficacy.
To build reliable precision medicine, we must understand why this confounding happens and how modern statistical and machine learning frameworks work to disentangle these two distinct biological signals.
The Two Dimensions of Patient Outcomes
As outlined by Shengri Tian et al. in their systematic review, Clinical implication of prognostic and predictive biomarkers for castration-resistant prostate cancer: a systematic review, biomarkers generally fall into two categories:
- Prognostic Signals: These reflect the natural history and intrinsic aggressiveness of the disease, regardless of the treatment administered. A prognostic marker tells you how a patient is likely to fare overall.
- Predictive Signals: These reflect a patient's likelihood of responding to a specific therapeutic intervention. A predictive marker identifies a biological mechanism that directly interacts with the drug's mode of action.
To see how easily these are confused, imagine a clinical study evaluating a new targeted cancer therapy. Patients with a specific genetic mutation show exceptionally long survival times after taking the drug. It is tempting to conclude that this mutation is a predictive biomarker for drug response.
However, if patients with this same mutation also live longer without the drug simply because their tumors grow more slowly, the mutation is actually a prognostic marker of low tumor aggressiveness. As Benjamin Haibe-Kains et al. noted in Predicting prognosis using molecular profiling in estrogen receptor-positive breast cancer treated with tamoxifen, failing to isolate these effects makes it impossible to determine whether a patient truly benefits from a specific drug or simply possesses a less aggressive disease phenotype.
Why Standard Machine Learning Models Fail
The confusion between prognostic and predictive signals is not just a data collection issue; it is baked into how standard machine learning models optimize their loss functions.
Consider a model trained to estimate the Conditional Average Treatment Effect (CATE). The goal is to predict the difference in an outcome $Y$ (such as survival time) if a patient receives a treatment ($W=1$) versus a control ($W=0$).
Mathematically, we can think of the outcome as being driven by two components: $$\text{Outcome} = f_{\text{prognostic}}(X) + W \cdot f_{\text{predictive}}(X)$$
Where $X$ represents the patient's baseline features. In almost all clinical scenarios, the baseline prognostic effect ($f_{\text{prognostic}}$)—the sheer variance in how sick patients are—is massive compared to the treatment interaction effect ($f_{\text{predictive}}$). Because standard loss functions minimize overall prediction error, a neural network or regression model will naturally devote its capacity to learning the dominant prognostic signal.
This leads to a severe failure mode when identifying feature importance. In a benchmarking study titled Benchmarking Heterogeneous Treatment Effect Models through the Lens of Interpretability, Jonathan Crabbé et al. demonstrated that models which appear highly accurate on paper often completely misattribute the cause of a treatment effect. The models achieve high accuracy simply by predicting that healthy patients stay healthy, mistakenly labeling general risk factors as the drivers of treatment effectiveness.
The Multi-Stage Learner Flaw
A common approach to CATE modeling is to use multi-stage or "T-learners," which train two independent models: one on the control group and one on the treated group.
As explained by David Svensson et al. in Overview and practical recommendations on using Shapley Values for identifying predictive biomarkers via CATE modeling, this architecture is highly susceptible to false positives. Because the two models do not share information, they may fit the dominant prognostic main effects in slightly different ways due to random noise or sample variance. When the algorithm subtracts the control prediction from the treated prediction to calculate the treatment effect, these minor algorithmic discrepancies are erroneously amplified and interpreted as true treatment interactions.
[Patient Features X] ───► [Control Model] ───► Predicted Baseline Outcome ┐
├──► [Difference] ──► False Treatment Interaction
[Patient Features X] ───► [Treated Model] ───► Predicted Treated Outcome ┘
The Clinical Confounding Trap
This statistical illusion mirrors a major bottleneck in clinical trial design. When researchers evaluate biomarkers using single-arm or non-randomized trials, they lack a proper control group to establish a baseline.
For instance, in studies analyzing imaging data or serum markers, such as the exploration of adipose tissue by Xin Huang et al. (Adipose tissue area as a predictor for the efficacy of apatinib in platinum-resistant ovarian cancer: an exploratory imaging biomarker analysis of the AEROC trial) or suPAR levels by Sven H. Loosen et al. (Serum Levels of Soluble Urokinase Plasminogen Activator Receptor Predict Tumor Response and Outcome to Immune Checkpoint Inhibitor Therapy), a marker may correlate perfectly with survival. Yet, without a comparative non-treated arm, it is impossible to prove whether the marker tracks a specific drug response or merely reflects a patient's robust baseline physiology.
Strategies for Disentanglement
To overcome this confounding, researchers use specific architectural constraints and validation strategies designed to force models to isolate interaction effects from main effects.
1. Explicit Treatment-Effect Architectures
Instead of letting a model freely learn any relationship, modern deep learning frameworks force a structural separation between the two signals. For example, Shuhan Xiao et al. developed a framework detailed in Enhancing Predictive Imaging Biomarker Discovery Through Treatment Effect Analysis that adapts CATE estimation directly for raw image data. By utilizing explicit statistical regularizers and Explainable AI (XAI) protocols, their approach prevents the network from relying on biased, handcrafted features, forcing it to isolate treatment-specific "predictive" image features from general "prognostic" ones.
2. Cross-Referencing External Control Datasets
When working with genomic or proteomic signatures, a candidate predictive model can be stress-tested against external datasets where patients did not receive the treatment. In Identification and Validation of Protein Biomarkers of Response to Neoadjuvant Platinum Chemotherapy in Muscle Invasive Urothelial Carcinoma, Alexander Baras et al. used this strategy to validate a two-protein signature. By proving that their signature correlated with survival only in the chemotherapy-treated cohort—and showed no correlation in untreated control cohorts—the authors successfully isolated a true predictive mechanism from general survival trends.
3. Unsupervised Invariance Learning
Another approach is to filter out the noisy, highly variable prognostic signals before modeling treatment effects. Qiyuan Li et al. demonstrated this in Consistent metagenes from cancer expression profiles yield agent specific predictors of chemotherapy response by developing an unsupervised method to extract "consistent metagenes." By identifying gene expression patterns that remain invariant across multiple independent datasets, they bypassed the common problem of overfitting to general disease severity, ultimately yielding robust, agent-specific predictors of chemotherapy response.
4. Interaction Modeling via Joint Frameworks
Rather than modeling treatment arms independently, researchers frequently employ joint statistical frameworks—such as Cox proportional hazards with explicit interaction terms or structured subgroup discovery algorithms like SIDES (Mattis Gottlow et al., Application of structured statistical analyses to identify a biomarker predictive of enhanced tralokinumab efficacy in phase III clinical trials for severe, uncontrolled asthma). These methods evaluate the marker and the treatment simultaneously, ensuring that a feature is only flagged as predictive if its mathematical interaction with the treatment is statistically significant, effectively neutralizing baseline prognostic noise.
Key Foundational Work
Predicting prognosis using molecular profiling in estrogen receptor-positive breast cancer treated with tamoxifen (2008) — contributed 2 concept elements to this direction.
Consistent metagenes from cancer expression profiles yield agent specific predictors of chemotherapy response (2011) — contributed 1 concept elements to this direction.