Routine lab trajectories flag cancer complications months ahead: 162 diagnoses, up to 6.1-fold enrichment on the rarest endpoint
Synopsis
Cancer treatment generates blood panels at nearly every visit, yet single-timepoint prognostic tools discard how those values move over time. A transformer trained on 2,777,595 laboratory measurements from 3905 patients with multiple myeloma or ovarian cancer predicted which of 162 treatment-associated complications would be newly recorded within two years, beating prevalence across all eight clinical categories with 1.5- to 6.1-fold enrichment. Across 91 internal endpoints with at least ten events, mean AUROC was 0.734 versus 0.710 for an engineered-trajectory baseline.
Interpretation
Routine laboratory sequences are treated as longitudinal physiological recordings, and the model outputs which specific complication will be newly recorded within two years rather than one aggregate prognosis score. The CARG and CRASH scores derive risk from pre-treatment values only, do not update as physiology changes under therapy, and do not resolve individual complication types; existing cancer laboratory-trajectory models predict only a small number of aggregate endpoints. All eight primary disease groups exceeded their respective prevalence within the 720-day window, with grouped AUROC from 0.65 for type 2 diabetes to 0.75 for fungal infections and enrichment of 1.5- to 6.1-fold, highest for myelodysplastic syndromes.
The model aggregates each visit day's laboratory values into one input token and lets a transformer learn joint changes across analytes over time. Inputs are restricted to 39 routinely obtained analytes, with missing values first completed by a separate transformer and then passed to the prediction model together with age and sex. Evaluation used a held-out test set of 390 patients and 1579 sequences; across 91 internal endpoints with at least ten events, mean AUROC was 0.734 for the transformer against 0.710 for the engineered-trajectory baseline, 0.692 for count-vector logistic regression and 0.664 for count-vector gradient boosting.
Predictive signal largely held across two mechanistically distinct cancers, indicating the model reads laboratory-observable physiology rather than cohort-level confounders alone. Five of eight grouped endpoints differed by no more than 0.07 in AUROC between the two subcohorts, including cardiovascular disease, bacterial infections, metastatic disease, type 2 diabetes and fungal infections. Evaluated separately on the myeloma subset (182 patients, 853 sequences) and the ovarian cancer subset (208 patients, 726 sequences); kidney disease and all-cause mortality diverged in opposite directions, and myelodysplastic syndromes diverged most sharply but rested on only 21 positive sequences from four patients.
Perspective
Evaluation covers 3905 patients with multiple myeloma or ovarian cancer at a single centre between 2000 and 2026, plus cross-system validation on two external cohorts, MIMIC-IV and MMRF CoMMpass; the authors state the retrospective study shows the signal is present and quantifies how far ahead it is detectable, but not that acting on it changes outcomes.
The authors report no prospective validation, and the model receives no treatment intensity, hospitalisation status or line of therapy, so observed associations may reflect disease progression, comorbidity burden or treatment effect; infection endpoints lose discrimination faster over longer horizons, and the 720-day horizon has limited clinical actionability for acute infectious complications; internal estimates for myelodysplastic syndromes rest on 21 positive sequences from four patients, which the authors call hypothesis-generating rather than a clinically mature use case.
