Autoencoder-Enabled Cox Regression for Metabolomics Survival Analysis
Predicting risk from thousands of lab values
Thousands of metabolite measurements come from the same samples a clinic already collects, and this work tests whether they improve risk prediction in chronic kidney disease beyond the handful of markers in use today.
Problem
Clinical risk models for chronic kidney disease run on a few routine markers: age, sex, kidney function, protein in the urine. Metabolomics panels measure thousands of small molecules from the same plasma and urine samples, and the hope is that they carry information those routine markers miss. Showing that they do is harder than it looks. With far more features than events and a model flexible enough to memorize its own training rows, an evaluation that looks careful can still report a gain that is pure optimism.
Approach
An autoencoder compresses metabolite profiles into a small latent space, and a Cox proportional hazards model is fitted on those latent dimensions together with the clinical covariates. Reconstruction, Cox, and latent-orthogonality objectives pull against each other during training, so the balance between them is learned rather than fixed by hand. The evaluation is the larger half of the work: a three-way split so the Cox model is never fitted on latents from the autoencoder’s own training rows, repeated cross-fitting across folds, and a selection protocol written and committed before the run rather than after seeing the numbers.
Outcome
In the cohort study the latent block adds a small but consistent improvement over the clinical baseline on all three endpoints: all-cause mortality, end-stage renal disease, and acute kidney injury. The added value replicates in every repeat. The analysis is still running and no manuscript has been submitted.