A machine learning model to predict risk of infection in patients receiving bispecific antibody therapy for multiple myeloma

Abstract

Abstract Background: Bispecific antibodies (bsAb) have revolutionized the care of patients with multiple myeloma (MM); however, they are also associated with an increased risk of infections, including infection-related mortality. There is an unmet need to identify clinical factors that evaluate the risk of infection among patients undergoing bsAb therapy, thereby allowing for earlier detection of patients at the highest and lowest infection risk, guiding tailored treatment decisions. We aim to develop a machine learning risk model to predict the likelihood of infection in patients receiving standard bsAb therapy. Methods: We used amulti-institutional cohort study of patients treated with at least one full dose of standard care teclistamab or talquetamab. Patient-level data, including disease and treatment characteristics and laboratory data, were retrospectively collected. Infections were identified and confirmed clinically, microbiologically, or by imaging and graded per CTCAE version 5. While infection prophylaxis was at per institutional practices, all patients received IVIG supplementation alongside bsAb. We developed a machine learning model using Python, applying AutoGluon-Tabular and PyTorch to develop a neural network, considering infection and severe infection (CTCAE Grade 3 and above) as binary classification problems, with a loss function optimizing for balanced accuracy. To avoid overfitting and address imbalanced data, we used k-fold bagging and automatic sample weighting and evaluated performance using out-of-fold predictions. Feature importance was assessed using SHAP (SHapley Additive exPlanations). Results: We included a total of 336 patients, 55% were male (n=185) and 45% were female (n=151). Approximately 72% of patients were Caucasian (n=243), 20% African American (n=67), and 8% (n=26) other. 260 patients (78%) had high-risk disease at diagnosis, and a total of 312 (93%) had triple-class refractory disease. A majority (82%, n=274) underwent prior autologous stem cell transplantation (ASCT) and 14% (n=47) underwent >1 transplant. Patients were followed for a median of 7 months, and teclistamab was used more often in our population (n=195 [58%] teclistamab, n=141 [42%] talquetamab). During the study period, 99 patients (30.6%) died. During the study period, patients with detailed dosing data available (n=301, 90%) received a median of 8 doses of bsAb (range 1-49). When considering infections occurring within the first year of starting bsAb therapy, a total of 139 patients (41%) developed any infection, with 84 (25%) experiencing a grade 3 or higher infection. A majority of those infections occurred within the first 90 days, with 103 (31%) developing any infection, and 58 (17%) developing a grade 3 or higher infection within 90 days. As previously noted in the literature, patients treated with teclistamab were more likely to develop infection as compared to talquetamab (n=98/195 [50.3%] vs. n=41/141 [29.1%], p=0.0002 by Chi-squared). Our neural network modeling successfully identified patients at risk of infection within 90 days of bsAb therapy, achieving an AUC of 0.69. The patients at highest risk of infection within 90 days were those who received teclistamab and received prior BCMA-targeted therapy, and those who experienced CRS (with increasing risk by grade). A risk reduction was noted for those who had undergone prior ASCT. Our modeling was also able to discriminate patients at risk of grade 3 or higher infection within 90 days of bsAb therapy with an ROC/AUC of 0.71, with significant contribution to risk by the above parameters, and the most significant risk contribution from the presence of extramedullary disease. Baseline laboratory data only modestly contributed to the risk of all grade and grade 3 and higher infections. When expanding to an ensemble model, we were able to markedly improve our accuracy, with an AUC of 0.91 for detecting grade 3 or higher infections within 90 days, at the risk of overfitting given our modest cohort size. Conclusions: To our knowledge, this is the first machine learning model capable of predicting the risk of any infection, including grade 3 or higher infections, within 90 days of initiating standard-of-care bsAb therapy for MM. Use of teclistamab therapy, prior BCMA-directed therapy, and CRS were associated with the highest risk of infection. Future research will focus on validating these findings in larger cohorts of patients treated with bsAb.