Associate Professor Roswell Park Comprehensive Cancer Center, United States
Introduction: The gold standard for monitoring response status in MM is the serum and urine protein electrophoresis (SPEP and UPEP) which quantify M-spike protein values. However, turnaround time for SPEP and UPEP is 3-7 days which delays treatment decisions.
Objective: To generate and validate a ML algorithm that predicts M-spike protein values in MM patients better than the currently available prognostic markers. Second, to demonstrate that ML technology is a transformative approach for public and population health that can expedite clinical decision-making.
Methods: Retrospective chart review was performed using de-identified, electronic medical records accumulated from January 2008 through January 2018. M-spike serum protein measurements (N=1,472 observations) were recorded at multiple time points for each patient. Median age of the cohort was 72 years (range: 42-96), and 45% were male. Random Forest (RF) analysis was performed to integrate 43 independent clinical and laboratory variables and generate a ML algorithm that predicted M-spike values as a continuous outcome variable.
171 patients diagnosed with active MM were included in the study. Forty-three independent readily-available demographic, laboratory chemistries, clinical values and SPEP values were measured by routine testing methods.
For RF analysis, patients were randomly split 50:50 into training and test sets based on per patient basis. 86 and 85 subjects were in the training set and test set, respectively, with 810 and 662 observations in each set. RF analysis identified the weighted value of each independent variable integrated into the ML algorithm. Pearson and Spearman correlation coefficients compared M-spike values determined using the ML algorithm and those determined using the gamma gap (GG) method with laboratory measured SPEP values. The prognostic M-spike value was measured by univariable and multivariable regression analyses.
Results: Pearson and Spearman correlation coefficients indicated that the ML algorithm correlated with laboratory measured SPEP values than values determined using the GG method. Feature selection modeling indicated that five input variables (serum IgG and IgM, first and second lagged M-spike values, and serum total protein) accurately predicted M-spike values (Pearson’s r = 0.96; RMSE 0.19). Feature selected modeling using only the first lagged M-spike value and serum total protein predicted the M-spike value (Pearson’s r = 0.95; RMSE 0.22).
Conclusions: Here we developed and validated a ML algorithm to reproducibly predict M-spike values using readily-available patient data. ML accurately predicts M-spike values better than conventional prognostic markers and represents a rapid, approach to leverage real-world patient information to longitudinally monitor disease burden. ML leverages vast amounts of real-world patient information to longitudinally monitor disease and can expedite clinical decision-making, reduce healthcare costs, and overcome disparities in healthcare access.
