Survival analysis—predicting the time until an event like a machine failure or a patient’s recovery—is a cornerstone of statistics. But in the real world, conditions change over time. A patient’s vitals fluctuate; a subscriber’s usage patterns shift. This "dynamic" setting is usually handled by a technique called landmarking, which essentially treats every new data point as a fresh, independent prediction problem.
In their NeurIPS 2022 paper, Temporally-Consistent Survival Analysis, Lucas Maystre and Daniel Russo argue we can do better. By borrowing a core concept from Reinforcement Learning (RL), they introduce a method that treats survival not just as a static regression problem, but as a recursive process that stays consistent over time.
The Problem with Landmarking
Traditional survival models often focus on a "static" snapshot, such as a patient's health at the start of a study. However, many modern datasets are "dynamic," consisting of sequences of states over time.
The current standard for dynamic data is landmarking: taking every intermediate state and regressing it against the final observed outcome. The problem with landmarking is that it ignores the relationship between consecutive states. It fails to recognize that your survival probability today is intrinsically linked to your expected survival probability tomorrow. This disconnect leads to high statistical noise and poor sample efficiency, especially when data is scarce.
Temporally-Consistent Survival Regression (TCSR)
To solve this, the authors propose Temporally-Consistent Survival Regression (TCSR). Its secret sauce is the temporal-consistency condition, inspired by the Bellman equations used in RL.
- The Recursive Idea: In a Markovian system, the probability of surviving $k$ steps from your current state is equal to the average probability of surviving $k-1$ steps from whatever state you transition to next.
- The Algorithm: Instead of regressing against the final "hard" outcome (e.g., did the event happen at time $T$?), TCSR regresses against a "soft" target: a combination of the immediate outcome and the model’s own prediction for the next state.
- Drop-in Compatibility: Crucially, TCSR can be implemented as a weighted binary classification problem. Researchers can use existing, optimized survival regression software and simply swap in the TCSR targets and weights.
Key Insights and Empirical Results
The paper highlights a beautiful symmetry: survival analysis is essentially "policy evaluation" where the "reward" is simply staying alive. TCSR is to survival regression what Least-Squares Temporal Difference (LSTD) is to standard RL regression.
Because TCSR links states together, it can "pool" information. If two different patients eventually reach the same health state, the model learns from both of them to refine its prediction for that specific state, effectively reducing noise. Furthermore, the model doesn't need to see a sequence finish to start learning; it learns from every single transition, making it highly robust for short, right-censored sequences.
The authors evaluated their approach across both real-world and synthetic datasets:
- Superior Sample Efficiency: On synthetic tasks modeling complex state transitions (such as a 20-dimensional Random Walk), TCSR reached the same level of performance as landmarking while requiring four times fewer samples. As seen in synthetic treatment effect estimations, TCSR maintains a significantly higher Kendall-$ au$ correlation with the true effect compared to landmarking, even as the number of treatments scales.
- Improved Clinical Accuracy: On the PBC2 dataset (tracking patients with a rare liver disease over multiple years) and the AIDS dataset (tracking CD4 cell counts), TCSR consistently outperformed both static baselines and landmarking in terms of the Concordance Index (CI)—a measure of how well the model ranks individuals by risk.
- Robustness: Despite being an iterative algorithm, TCSR proved to be very stable in practice, often converging in fewer iterations than the number of time steps ($K$) in the data.
Limitations and Future Work
While powerful, TCSR does have boundaries. The theoretical guarantees rely on the system being a Markov chain (where the future depends only on the current state). While feature engineering can mitigate this, highly non-Markovian data remains a challenge.
Additionally, the current framework operates in discrete time steps. Extending this to continuous-time, irregularly sampled data is a promising next step. Finally, the paper currently models a single type of event. In many real-world scenarios, there are "competing risks" (e.g., a patient might leave a study or experience an unrelated event), which the model would need to be extended to handle.
Summary updated