Prognosis in multiple myeloma, a cancer of plasma cells in the bone marrow, still rests largely on staging at the moment of diagnosis, discarding the stream of information generated as treatment unfolds. Researchers at the University of Tokyo's Graduate School of Frontier Sciences, working with colleagues at RIKEN, have built a deep learning model designed to use that neglected data to forecast a patient's remaining overall survival as their disease evolves.
The model fuses three kinds of input: gene expression profiles from diagnosis, converted into images using the group's DeepInsight technique; clinical and blood test results taken at diagnosis; and longitudinal changes in ten laboratory test items together with treatment history. Drawing on observation windows of one to eighteen months after diagnosis, it produced accurate survival predictions from the point of analysis onwards.
Published in Briefings in Bioinformatics, the approach points towards genuinely dynamic prognosis, in which risk estimates are updated as therapy progresses rather than fixed on day one. The authors suggest the framework could extend to other long-term diseases where treatment generates rich time-series data, offering clinicians a continuously refreshed picture of how an individual patient is likely to fare.