The Cost of Binarizing Survival Outcomes in Clinical Prognostic Modeling

In clinical machine learning, survival analysis is the gold standard. It predicts time-to-event outcomes, such as how long a patient survives after a cancer diagnosis. However, many researchers still simplify this complex temporal data. They convert it into a binary "yes/no" target. For example, they might ask, "did the patient survive 2 years?" This practice treats life as a coin flip at a fixed deadline. It discards everything that happens before or after that point.

While this simplification makes modeling easier, it creates a massive blind spot. Collapsing a continuous timeline into a single threshold loses information. You cannot see how certain factors influence the speed of an event. This paper argues that binarization is not a neutral choice. It is a destructive one. It actively hides important biological signals. The authors introduce the Survival-Aware Bayesian Network (SA-BN). They aim to prove that treating survival as a time-to-event process recovers the features clinicians rely on most.

The signal lost to fixed horizons

Current Bayesian network (BN) modeling for cancer often discretizes survival into a binary indicator. This approach is widespread. Standard BN structure-learning algorithms—which determine how variables relate to one another in a graph—typically require discrete targets. The logic is often one of methodological convenience. If you want to know if a patient survives 2 years, you just label them 0 or 1.

But this creates two distinct failure modes. First, binarization excludes censored patients. These are patients who have not experienced the event by the end of the study. This means you throw away valuable data. Second, it masks "graded effects." Many clinical markers do not act like light switches. They act like dimmers. A heavy smoker does not just "fail" a 2-year survival test. They tend to die sooner than a moderate smoker. When you binarize the outcome, that continuous gradient disappears. The model becomes unable to recognize the feature as a meaningful predictor. As seen in the comparison of Markov blankets in the radiation cohort [Table 2], the binary approach fails to pick up these nuanced drivers.

Decoupling structure from the clock

The authors solve this by proposing the Survival-Aware Bayesian network (SA-BN). This is a hybrid architecture. It separates how features relate to each other from how they relate to the survival outcome. The core innovation is a decoupling of the scoring functions used during structure learning.

The pipeline works in four main stages:

1. Feature-Feature Scoring: The model uses standard discrete scoring functions (like BDeu or K2) to learn dependencies between clinical variables. This part of the graph remains unchanged from traditional BNs.
2. Feature-Target Scoring: Instead of using a classification score for the link between a feature and the survival node, the authors use the Cox partial log-likelihood. This is a mathematical formulation that accounts for both the timing of the event and the presence of censored patients.
3. Likelihood-Ratio Testing: Edges between features and the survival target are added only if they pass a likelihood-ratio test (LRT) with $p < 0.05$. This ensures that only statistically significant temporal relationships are included.
4. Consensus Discovery: To ensure the graph is not an artifact of a specific algorithm, the authors identify the "Markov blanket." This is the minimal set of variables that makes the target conditionally independent of the rest of the network. They find this by taking a consensus across three different structural learning methods.

The resulting graph, shown in for the radiation cohort, identifies a survival-aware Markov blanket.

It includes critical continuous variables that the binary model ignored.

Recovering the hidden gradients

The empirical evidence suggests the "cost" of binarization is significant. In the radiation cohort, SA-BN recovers 9 features compared to the 8 found by the binary proxy. The qualitative difference is more important. Specifically, SA-BN recovers smoking pack-years and overall cancer stage. Both have continuous, graded effects on survival.

The paper reports several key performance changes: * Concordance Index (C-index): This metric measures how well the model ranks survival times. In the surgical cohort, moving from a binary Markov blanket to an SA-BN blanket improved the test-set C-index from 0.648 to 0.726. * Binary Classification Gain: Using features selected by the survival-aware model improves even simple binary tasks. In the radiation cohort, the AUC-ROC (Area Under the Receiver Operating Characteristic curve, a measure of classification accuracy) rose from 0.797 to 0.826. * Causal Interpretability: SA-BN maintains the temporal structure. This allows the authors to use G-computation to estimate the "interventional RMST" (Restricted Mean Survival Time). This expresses the effect of a feature in actual months of life gained or lost, as shown in and .

The authors ran an ablation study to test the cause of this improvement. They wanted to see if it came from just having more data. SA-BN does include censored patients. However, they found that even when running SA-BN on the exact same subset of patients used by the binary model, the Cox scoring formulation was the primary driver [Table 3].

Measuring the edges of the method

There are clear boundaries to where SA-BN excels. The advantage is most pronounced when datasets contain features with strong graded effects and substantial censoring. If your features are all binary (e.g., "received treatment" vs. "did not receive treatment"), the delta will shrink.

There are also two technical caveats: 1. Hybrid Complexity: SA-BN is a "hybrid" method. Feature-target edges use the Cox likelihood. However, feature-feature edges still rely on discretization. The authors note that a fully survival-native scoring function for the entire graph would be more theoretically coherent. 2. Small Sample Sensitivity: The surgical cohort test set contained only 35 events. While the direction of improvement was consistent, the bootstrap confidence intervals for the C-index overlapped. This means the magnitude of the gain in smaller cohorts requires more validation.

The verdict: Stop simplifying your targets

If you are building prognostic models for clinical use, the verdict is clear: stop binarizing your survival endpoints.

Converting time-to-event data into a binary classification task is often done for algorithmic convenience. However, this paper demonstrates a fundamental loss of biological truth. By ignoring the temporal gradient, you blind your model to important prognostic drivers. These include hemoglobin, smoking history, and inflammatory markers.

The SA-BN approach is a relatively low-overhead modification. It is essentially a drop-in replacement for the scoring step in a Bayesian network pipeline. It ensures that your feature selection is actually looking at the right signal. If you want to move beyond simple "will they or won't they" predictions, you must treat survival as a timeline, not a threshold.

Code for the implementation is reportedly available at https://github.com/xinformatics/COBS.

Figures from the paper