Imagine evaluating two cancer patients undergoing different treatments. Patient A chooses an aggressive surgery: the immediate risk of mortality is exceptionally high due to surgical complications, but if they survive the first month, their long-term prognosis is excellent. Patient B chooses a conservative drug therapy: the immediate risk is negligible, but the disease progresses steadily over the years.
If you plot the survival probabilities of these two patients over time, their curves will inevitably cross.
Traditional survival analysis, anchored by the classical Cox Proportional Hazards (CoxPH) model, is fundamentally incapable of modeling this scenario. It relies on the proportional hazards assumption, which dictates that the relative risk (the hazard ratio) between any two individuals must remain constant over time. If Patient A is twice as likely to die as Patient B on day one, they must be twice as likely to die on day one thousand.
As noted by Nagpal et al., this rigid constraint fails to capture the non-linear interactions and time-dependent risks inherent in real-world clinical data. To accurately map diverse disease trajectories, modern survival modeling has shifted toward frameworks that relax or entirely eliminate this assumption.
The Mathematical Trap of Proportionality
In a standard proportional hazards framework, the hazard function $\lambda(t|x)$ for an individual with covariates $x$ is modeled as:
$$\lambda(t|x) = \lambda_0(t) \cdot \exp(\beta^T x)$$
where $\lambda_0(t)$ is a baseline hazard function over time, and $\exp(\beta^T x)$ is a static risk multiplier. Because the time component $t$ is completely isolated from the patient features $x$, the ratio of hazards between two individuals simplifies to a constant that never changes with time.
When real-world data violates this assumption, the mathematical consequences are severe. For instance, Jiang et al. demonstrate that non-proportional hazards render hazard ratios "non-transitive." This means that if Treatment A appears superior to Treatment B, and B superior to C in separate trials, the mathematical relationship breaks down, potentially leading indirect treatment comparisons to yield contradictory or false conclusions.
To bypass this trap without losing statistical validity, researchers often turn to alternative metrics. Gao et al. advocate for Restricted Mean Survival Time (RMST). Instead of calculating a single, flawed risk multiplier, RMST measures the area under the survival curve up to a specific time horizon. This captures the cumulative treatment effect regardless of whether the risk fluctuated or the survival curves crossed along the way.
Survival Probability
1.0 | \ Patient B (Drug)
| \_______
| / \____
| / \___ Patient A (Surgery)
0.0 |____/____________________
0 Time
^ Crossing Point (Non-Proportional Hazard)
Discretizing Time into Dependent Tasks
How do we design machine learning architectures that naturally allow survival curves to cross? One intuitive approach is to transform the continuous time-to-event problem into a sequence of discrete intervals, treating each interval as a distinct prediction task.
Yu et al. pioneered this approach with Multi-task Logistic Regression (MTLR). Instead of estimating a global hazard ratio, MTLR models the probability that an individual survives past a series of discrete time points $t_1, t_2, \dots, t_m$. The model optimizes these predictions simultaneously as a sequence of dependent regressors.
By parameterizing each time step independently while enforcing logical consistency (e.g., a patient cannot be alive at $t_2$ if they died at $t_1$), MTLR allows the predictive weight of each clinical feature to vary dynamically across the timeline. A biomarker might have a massive coefficient in the first month but drop to zero by year five, allowing individual survival curves to take completely unique, intersecting shapes.
Inductive Biases via Neural Frailty
While deep neural networks can theoretically learn any non-linear function, standard architectures often struggle to capture the specific temporal dynamics of survival data without explicit constraints. To address this, researchers have reintroduced classical statistical concepts as inductive biases within deep learning models.
One powerful mechanism is frailty—an unobserved, multiplicative random effect that accounts for hidden heterogeneity within a population. Wu et al. introduced the Neural Frailty Machine (NFM) to exploit this property.
Mathematically, if an individual's hazard is modified by an unobserved frailty term $Z$, the population-level hazard ratio naturally decays over time, even if the underlying individual hazards are perfectly proportional. This happens because high-risk individuals experience events earlier, leaving a healthier, lower-risk subpopulation behind as time progresses. By integrating a frailty term directly into a neural network's loss function, the NFM naturally models crossing hazards and non-proportional risk patterns without requiring explicit time-varying coefficients.
Continuous Time and Dynamic Covariates
In many clinical and industrial applications, risk factors are not static baselines measured once at the beginning of a study; they are continuously moving targets. Vitals spike, blood panel results fluctuate, and environmental conditions evolve.
To model these continuous interactions without falling into the proportional hazards trap, Wang et al. developed BoXHED (Boosted eXact Hazard Estimator with Dynamic covariates). Instead of relying on a pre-defined functional form, BoXHED uses tree-boosting to estimate the hazard function directly from time-dependent covariates.
By calculating the exact hazard at any given timestamp as a function of the current, fluctuating covariate state, the model completely bypasses the need for a static hazard ratio. This allows the model to capture real-world scenarios where the immediate risk of an event is driven entirely by the real-time volatility of the underlying features rather than a historical baseline.