ASH 2024 | Development and validation of novel machine learning-based risk scores for multiple myeloma

So this is a project that we have developed in collaboration with the HARMONY Big Data Consortium in Europe. So the consortium has gathered data from a lot of clinical trials which have been undertaken by collaborative groups in Europe for the past 20 years or so. And the idea was, can we actually analyze this data with machine learning to develop individualized risk scores and compare these risk scores with standard risk classification systems like the revised ISS or the revised-2 ISS. And for that, we ran a random forest algorithm, which is a supervised machine learning system for predicting survival and both overall survival and PFS, so globally to predict risk. And we ended up with three different machine learning models, which are the following ones. The first one is based on clinical and cytogenetic data. This model includes 1q gain and 17p deletion in the model. And the accuracy is quite good, it’s 0.67 in C index, which is quite high. The second model does the same, but without cytogenetics. So its accuracy is a little bit lower, but it has been compensated through machine learning with the impact of other variables. And this model is particularly oriented towards risk stratification in areas where cytogenetics and molecular genomics are not available, like Latin America, for example. And the third model, it’s quite novel, by the way, because it’s a dynamic risk model. So what we did was to take the original clinical plus cytogenetics model and recalibrate risks with best response achieved to induction therapy. So we observed that by doing this, we were actually increasing our accuracy a lot, and it actually surpassed 0.7 in C-index. And we can obviously better refine the risk of our patients in the future based on how they have responded to the upfront therapy. And we have finally developed an online calculator with the HARMONY Consortium where the clinicians can actually calculate the scores for their patients and see the information structure, whether the patient is transplant eligible or not transplant eligible and whether the patient was intended to be treated with an IMiD, PI, or IMiD-PI combined therapy. Sadly, we did not have patients with daratumumab in first line, so we do not have the data on this still, but probably we’ll do additional collaborations in the future to test how the models perform in these patients.

This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.