Artificial intelligence-enabled screening strategy for drug repurposing in monoclonal gammopathy of undetermined significance

In this manuscript, we describe a method for generating drug repurposing hypotheses in MGUS using EHR data and explainable machine-learning. We accomplished this using an XGBoost Cox survival model and Shapley feature explanations. Our model achieved an adequate fit of the survival data.

This study represents the first application of machine-learning for screening drug repurposing candidates in MGUS. We propose this methodology as a low-cost precursor prior to examining drug candidates of interest in synthetic clinical trials or prospective trials. Completing this type of study from EHR data of course requires access to a relatively large cohort of patients with a given disease and fairly complete follow-up data, both of which were available due to our status as an academic tertiary referral center.

When examining feature hazard ratios shown in Table 3, several medication classes appeared to be associated with reduced odds of MGUS progression, specifically, multivitamins, immunosuppression, non-coronary NSAIDS, proton pump inhibitors, vitamin D supplementation, opioids, statins, and beta-blockers. The literature review did not reveal any known associations between multivitamin, NSAID, or opioid use and MGUS progression. Regarding immunosuppression, tacrolimus, cyclosporine, and methotrexate were included. The literature review suggested that in both renal transplant and liver transplant populations, there was no association between the tacrolimus versus cyclosporine-based immunosuppression regimens and the development of MGUS or MGUS progression outcomes, though the number of such outcomes was small [7,8,9]. No studies reported an association between methotrexate use and MGUS progression risk. Overall, it is possible that this protective association with immunosuppression may reflect the fact that patients on these drugs typically undergo extensive laboratory evaluations for their comorbidities, and thus a greater number of benign MGUS cases are detected relative to the broader population.

With respect to proton pump inhibitors, our analysis also suggested a significant protective association with proton pump inhibitors as well, a finding that has not previously been reported. There have, however, been reports of progressive intestinal microbiome disturbances in patients with MGUS and multiple myeloma, compared to normal patients [10]. Proton pump inhibitors are known to modify the intestinal microbiome, in ways generally thought to be deleterious; however, our finding of a positive association with MGUS outcomes may warrant further investigation [11]. With respect to Vitamin D supplementation use, prior research has demonstrated significantly lower levels of vitamin D2 and provitamin D3 in Waldenström Macroglobulinemia patients relative to IgM MGUS patients [12]. A causative role has not yet been established, however.

Limited research has been devoted to any association between statins and MGUS progression, but one letter to the editor reported no relationship between statins and MGUS progression in a 200-patient case-control study, while another study reported an association between statin use and improved multiple myeloma survival in a cohort study [13, 14]. Additionally, a network meta-analysis has demonstrated an all-cause mortality benefit with statin use among patients with or at risk for cardiovascular disease [15]. With respect to beta-blockers, a prior retrospective cohort study noted better outcomes among multiple myeloma patients on beta-blockers; similar results in MGUS patients have not previously been reported. Regarding thyroid supplementation, at least one population-based study, which included 19,303 patients with MGUS, noted a lower risk of disease progression in patients with autoimmune disease; the association between lower progression risk and the presence of thyroid supplementation may reflect this [16]. Overall, it is notable that, while associations between MGUS outcomes and metformin use appear to have been the most deeply explored in the literature to date [17,18,19], our primary analysis notes a stronger association with multivitamins immunosuppression, proton pump inhibitors, NSAIDS, Vitamin D supplementation, opioids, statins, and beta-blockers. Additionally, our sensitivity analysis limited to those with high M spike suggested that multivitamins, non-coronary NSAIDs, and metformin were associated with lower odds of MGUS progression. Analogously, our sensitivity analysis excluding those with IgM MGUS suggested that multivitamins, immunosuppression, non-coronary NSAIDs, statins, PPIs, and opioids were associated with lower odds of MGUS progression, while loop diuretics were associated with higher odds of progression.

Regarding non-medication feature hazard ratios, we note that body mass index and serum M-spike levels were associated with significantly higher odds of MGUS progression, consistent with existing literature [20].

Limitations

Our study was limited by the inherent shortcomings of using her data to ascertain outcomes and medications. Detection of disease progression outcomes may be incomplete. We also note that our MGUS database may contain patients with small M spikes associated with autoimmune diseases, even though these may not represent clones with the potential to progress to malignant disease. We attempted to mitigate this through sensitivity, analysis, however. Medication data also may be incomplete, and we were unable to ascertain medication adherence or prescription durations. This may have contributed to the large proportion of patients taking corticosteroids, which may have been for short durations in some cases. In future studies, medication data could be enriched by linking to claims data. In our study, we were unable to obtain medication data for 27% of patients either because the medication history was not entered into the EHR or because the patients were not taking any medications. Demographic and laboratory characteristics between the two groups, nonetheless, were similar. As expected, the no-medication group had significantly fewer medical comorbidities. Finally, our study is also limited by its retrospective nature, limiting any inferences about causality.

In conclusion, we analyzed EHR data on a large cohort of MGUS patients using explainable machine-learning to examine associations between patients’ medications and their MGUS outcomes. We uncovered associations that have been previously suggested (better hematologic malignancy outcomes with statin and beta-blocker use) and others that have not (decreased risk of progression with proton pump inhibitor use). Of note, we detected the strongest associations among drugs that have received relatively less attention in related literature to date, namely, multivitamins, immunosuppression, proton pump inhibitors, and vitamin D supplementation. Future research should focus on prospectively investigating these associations and applying similar methodology to other disease states.

留言 (0)

沒有登入
gif