Jerina Hoxha
30.4.25
Machine learning model anticipates disease trajectories to support early intervention and personalised care
Our latest publication in npj Digital Medicine, "Predicting progression events in multiple myeloma from routine blood work", presents a modular artificial intelligence model that predicts disease progression in multiple myeloma using routine lab data. Led by CERTAINTY partner Maximilian Ferle and collaborators from Leipzig University, Fraunhofer IZI, and Memorial Sloan Kettering, the study introduces a tool designed to forecast changes in blood parameters and flag progression events before they are clinically apparent.
Background
Multiple myeloma is characterised by complex disease kinetics and variable outcomes, which current staging systems often fail to capture beyond the initial diagnosis. Leveraging the wealth of data generated through routine monitoring, the authors trained a neural network to learn individual patient trajectories and anticipate future developments.
Key Findings
Accurate short-term forecasting and robust progression detection
The model outperformed traditional methods in predicting future values for key lab parameters, including M-protein, serum free light chains, and haemoglobin. When combined with a progression annotation module, it reached a sensitivity of 92 percent and AUROC of 0.88 for progression detection.
External validation confirms generalisability
Applying the model to an independent dataset from the GMMG-MM5 study confirmed its performance, with AUROC values of 0.87 and a predictive value significantly above chance even in low-prevalence scenarios.
Supports digital twin vision for haematological care
Designed as a modular and interpretable system, the model aligns with digital twin principles. It enables dynamic risk assessment, potentially reducing clinical progression through earlier detection and more frequent follow-up when needed.
Conclusion
This study demonstrates the potential of AI in real-world oncology, using only standard laboratory measurements to provide individualised forecasts. By enabling earlier intervention without costly diagnostics, it represents a step forward in realising scalable and patient-centred digital health tools.
Access the Publication
Read the article in npj Digital Medicine: https://doi.org/10.1038/s41746-025-01636-9