Routine laboratory trajectories encode the onset of organ-level complications in cancer

Technical Summary: Routine Laboratory Trajectories Encode the Onset of Organ-Level Complications in Cancer

Problem Statement
Patients with advanced cancer, particularly those requiring prolonged multi-line therapy such as multiple myeloma and ovarian cancer, progressively accumulate organ-level complications across multiple systems. While routine laboratory panels are drawn at nearly every clinical encounter, generating dense longitudinal records of organ function, this temporal signal remains largely underexploited. Existing prognostic tools, such as the CARG score or CRASH scale, rely on pre-treatment values and fail to update as physiology changes or incorporate laboratory trajectories across treatment lines. Furthermore, current disease-specific models often predict aggregate adverse events or survival rather than the onset of individual, treatment-associated complications. There is a lack of models that apply raw laboratory trajectories to complication-resolved risk estimation within defined cancer populations using standard, routinely collected data.

Methodology
The study trained a transformer-based model to predict the two-year onset of 162 treatment-associated complications across eight clinical categories (including myelodysplastic syndromes, cardiovascular disease, kidney disease, and infections).

• Data Source and Cohort: The model was trained on 2,777,595 structured laboratory measurements from 3,905 patients (1,855 with multiple myeloma, 2,050 with ovarian cancer) treated at a single tertiary center between 2000 and 2026. The dataset was filtered to include only 39 routinely tested analytes (spanning hematology, metabolic, hepatic, inflammatory, coagulation, and endocrine panels) to avoid indication bias from specialized testing.
• Data Processing: Laboratory trajectories were segmented into non-overlapping observation windows of 10–30 days. Missing values were handled by a dedicated transformer-based imputation model trained to reconstruct masked analytes, ensuring the prediction model relied on physiological values rather than testing patterns.
• Model Architecture: A bidirectional transformer encoder (based on Qwen-3 architecture components, including rotary positional embeddings and SwiGLU activations) was trained from scratch on structured laboratory data without language pretraining. Input tokens encoded z-score-normalized values of the 39 analytes and patient age, with a binary sex token appended per sequence.
• Task Formulation: The task was framed as multi-label classification to predict the probability of each of the 162 diagnoses (mapped from ICD-10 codes) occurring within 730 days (two years) after the input window.
• Baselines and Validation: The transformer was compared against non-sequential baselines (XGBoost and logistic regression) trained on bag-of-words representations of the same data. External validation was performed on two independent US cohorts: the MIMIC-IV critical care database and the MMRF CoMMpass multiple myeloma registry, without fine-tuning or cohort-specific recalibration.

Key Results

• Predictive Performance: The transformer predicted the two-year onset of all eight primary disease groups above their respective prevalence baselines. AUROC values ranged from 0.65 for type 2 diabetes to 0.75 for fungal infections. The model achieved significant predictive enrichment, with the highest ratios observed for rare endpoints (e.g., myelodysplastic syndromes showed a 6.1-fold enrichment; fungal infections, 3.1-fold).
• Temporal Advantage: The transformer matched or outperformed non-sequential baselines across all eight grouped endpoints, with AUROC gains up to +0.11 for myelodysplastic syndromes and +0.09 for fungal infections. This demonstrates that longitudinal trajectories capture evolving, complication-specific physiology inaccessible from isolated measurements.
• Biological Coherence: Predictions generalized across both cancer types for five of eight endpoints (cardiovascular, bacterial infections, metastatic disease, type 2 diabetes, fungal infections), despite mechanistic differences in disease pathophysiology. Divergence was observed in endpoints where the laboratory signature is disease-specific (e.g., myelodysplastic syndromes, where the ovarian cohort had zero prevalence in the internal test set).
• Biomarker Signatures: Masking analyses revealed that the model relied on interpretable biomarkers consistent with known pathophysiology. For example, glucose dominated type 2 diabetes predictions, glomerular filtration estimates (eGFR) drove kidney disease predictions, and red blood cell counts were critical for myelodysplastic syndromes. Pairwise masking indicated synergistic relationships between biomarkers.
• External Transferability: The model demonstrated transferability to independent healthcare systems (MIMIC-IV and MMRF CoMMpass) without retraining. Renal and metabolic endpoints transferred most robustly (AUROC up to 0.85 in MIMIC-IV for specific kidney disease stages), while performance for infectious complications varied, likely due to coding system differences.

Significance and Claims
The authors claim that routine oncological laboratory data encode organ deterioration weeks to months before clinical onset. The primary significance of this work is the demonstration that a "complication-specific" signal exists within standard, routinely collected data, enabling surveillance without additional testing infrastructure.

The paper posits that this approach offers a "cheapest dense longitudinal monitoring channel available in oncology," as it utilizes data already produced by every oncology unit and stored in hospital information systems. By resolving 162 specific endpoints rather than aggregate risks, the model provides actionable, complication-specific guidance that current models lack. The authors emphasize that the predictions are driven by interpretable biomarker signatures aligned with established pathophysiology, keeping the model auditable rather than opaque.

The study concludes that while prospective validation is required to establish decision thresholds for clinical intervention, the data infrastructure for detecting these temporal signals is already in place, offering a pathway to patient-level digital twins of cancer where organ-state representations are continuously updated.