Cancer side effects can be forecast from routine labs: SPARC holds up across 11 hospitals
Synopsis
Predicting cancer treatment side effects has long relied on static nomograms built from a handful of variables. SPARC instead models longitudinal sequences of blood counts, metabolic panels and vital signs, pretraining on 101,920,379 measurements from 224,047 patients and then fine-tuning for mortality, anemia, neutropenia, febrile neutropenia, acute kidney injury and liver toxicity. It matched or led comparison models on held-out, prospective and multi-hospital cohorts, and its predicted risk strata identified patients most likely to benefit from G-CSF prophylaxis.
Interpretation
The prediction target moves from a single snapshot to each patient's own trajectory of labs and vital signs. Earlier cancer mortality and adverse event prediction generally relied on static nomograms built on a small number of variables and limited to specific disease and treatment contexts. On the held-out MSK test set, SPARC outperformed or matched all comparison models on both mortality tasks and on acute kidney injury, anemia, liver toxicity and neutropenia; only the rare febrile neutropenia task was best predicted by TabPFN.
The model splits each visit into clinical modules and encodes how much time separates two measurements. SPARC is a time-series transformer foundation model pretrained with a missingness-aware autoregressive objective, predicting the last timepoint of a window from earlier ones, then fine-tuned per task. Pretraining used 157K windows averaging 10.2 timepoints and 211.9 measurements; held-out future lab values were predicted with mean squared error 0.423, versus 0.872 for the feature mean, 0.590 for carrying forward the most recent value, 0.519 for per-feature linear regression, 0.487 for ridge regression on most-recent values and 0.451 for ridge regression on historical values.
Pooled fine-tuning across hospitals beats the zero-shot baseline in more task-site pairs than tuning each site alone. Zero-shot results at the six INSIGHT hospitals held for higher-prevalence tasks, while febrile neutropenia and neutropenia proved harder to predict. Pooled fine-tuning improved on the zero-shot baseline in 24 of 30 task-site pairs versus 15 of 30 for site-specific fine-tuning, with 10 versus 1 discordant pairs and McNemar exact p = 0.0117.
Perspective
The validated scope is patients receiving cancer treatment at academic medical centers in large US cities, with model inputs limited to complete blood counts, comprehensive metabolic panels and routine vital signs; the authors note that laboratory testing patterns may differ in community and non-US hospitals. SPARC is released as a community resource, and the authors say the results motivate interventional randomized studies of G-CSF administration.
In the prospective cohort, the only task whose AUROC consistently degraded by more than 5% was febrile neutropenia, which the authors attribute to the small number of events and changes in patient case mix over time. At CAIA external centers, febrile neutropenia was again the only task that consistently degraded, by at most about 10% in AUROC, which the authors attribute to limited training data, lower metric precision for rare events and institutional differences. Acute kidney injury was predicted with AUROC 0.87 in patients receiving bisphosphonates and 0.64 in patients receiving cisplatin, which the authors link to clinicians being less inclined to prescribe cisplatin to patients with underlying renal dysfunction. The authors also report that performance on most tasks plateaued with one-eighth of the pooled data, while liver injury kept improving.