Survival Analysis
Ataraxis advances survival analysis with multimodal models that integrate pathology-derived features with clinical data, estimating how a patient's risk evolves over time rather than producing a single static score. Our models deliver calibrated probabilities of recurrence or survival at different time points that support personalized treatment planning and form the foundation for modeling treatment effects.
Why survival analysis matters
Classification models make categorical predictions, such as “high risk” or “low risk”. Survival analysis goes further, estimating the probability of recurrence across specific time horizons such as within 5 years, 10 years, or beyond. A 20% chance of recurrence within 2 years represents a very different clinical situation than a 20% chance over 10 years, and the difference matters for treatment decisions. By accounting for incomplete data, survival analysis provides a richer, more clinically meaningful view of prognosis that better guides treatment intensity, surveillance, and follow-up.
Our models are trained on diverse real-world datasets spanning multiple institutions and cancer types. Pathology and clinical signals are trained together so the system learns jointly from both modalities, using time-to-event models.
By combining the interpretability of survival analysis with the representational power of foundation models, we transform prognosis from a fixed snapshot into a dynamic, patient-specific prediction that adapts to individual biology and clinical context.
We perform extensive hyperparameter tuning to build a pool of high-performing candidate models, then select the top performers and aggregate them into an ensemble. Averaging across these complementary models reduces variance, improves generalization, and produces more robust and stable patient risk estimates than any single model alone.
Legacy prognostic tests report relative risk scores that rank patients against one another but are not designed to reflect the actual predicted probability of recurrence. Our models are explicitly calibrated so that predicted risks align closely with observed outcomes i.e., the model's 15% five-year recurrence risk aligns closely with absolute real-world risk. This calibration holds across independent test sets spanning diverse patient populations, making our risk estimates not just discriminative but clinically actionable in an absolute sense.