Prognosis and Therapeutic Response Transfer Learning Model for Multiple Myeloma

The combination of clinical knowledge (applicable for a majority of patients, published in the form of review articles), real world evidence (describing more nuanced outcomes for small cohorts) and innovative artificial intelligence algorithms opens potent avenues to re-examine clinical findings and uncover new biomarkers for the prognosis / prediction of therapeutic responses--in a manner that can directly be incorporated in clinical decision support tools.

In this work, neural networks are first developed to reproduce clinical guidelines with >95% accuracy. After mastering the complex knowledge that is generally expected from human doctors, a transfer learning technique was used to sift through de-identified longitudinal data of 9267 patients with multiple myeloma-related conditions at the Vanderbilt University Medical Center using progression free survival (PFS) to quantify therapeutic outcomes. The "precision medicine neural networks" obtained as a result can be compared with conventional and less portable survival model algorithms, using Shapley values to explain prediction differences.

Testing a first hypothesis that "lytic bone lesions are a prognostic factor for poor PFS", a study involving 1530 patients confirmed that the median PFS of 54 ± 6 months (684 censored patients showing progression) in the presence of bone lesions is significantly lower than the 107 ± 40 months (90 censored patients) in the absence of lesions.

Keeping only 179 patients for which a full range of cytogenetic factors are available and using a Cox regression & Random Survival Forests that provide the best fit of the data, we confirmed previous findings for high-risk trisomies 1 and 7, monosomy 13, deletion 12p and translocation t(11;14)(q13;q32). We furthermore uncovered new additional adverse factors del6q, 2+, 21-, 16- to formulate a model that achieves a statistically significant concordance of 0.72 ± 0.07.

Comparing therapeutic effects for patients in a real-world hospital setting with clinical trials, we found that lenalidomide + bortezomib + dexamethasone used in a first-line therapy resulted in lower median PFS of 17-47 months than the 39-52 months published in the SWOG S0777 trial, most likely because comorbidities contributed to shortening the PFS in real-world settings.

We conclude with an analysis of concrete examples where therapeutic recommendations differ from guidelines, explaining the reason with statistically significant cohorts observed in the data.