ASH 2024 | Validation of a machine-learning system to predict the diagnosis of B-cell lymphomas

Yeah, okay, so we have a very big program in our labs, at least in Europe, where I think it’s worldwide. So we have a lot of flow cytometry resources to precisely diagnose neoplasms, particularly hematological neoplasms, but the demand is growing so fast that the personnel to do this is so limited that we have to find an equilibrium. And our idea as clinicians doing also machine learning and doing diagnosis was to develop an ad hoc system for our facility that could work as a virtual assistant and could analyze automatically flow cytometry data from patients with a suspicion of non-Hodgkin lymphoma, CLL, monoclonal B-cell lymphocytosis, diffuse large B-cell lymphoma, follicular lymphoma, mantle cell lymphoma, T-cell lymphoma, whatever lymphoma. And be capable of comparing this with our training dataset, which were our patients from the past four years who were annotated by humans, human experts, and provide us feedback on how likely that sample is actually pathogenic, how likely is that sample a lymphoma sample. So we have developed this project with 3,500 samples in the training set of which one third was actually lymphoma and two-thirds were not lymphoma and we then ran like a supervised clusterization of the cells followed by a supervised machine learning system whose endpoint was to predict the diagnosis made by the human and our results are quite good particularly for B-cell lymphoma where we achieved enhanced validation with of 95 a rock AUC and that’s quite robust. And interestingly, we observed that in the patients who were misclassified by our system, or at least where we find a disparity between the prediction of the system and the results reported by the human, there were some cases where our system was failing, but in other cases we observed that there was a human failure and there was a diagnosis not being made by a human because, you know, it’s normal. we make some great mistakes. And finally, we have collapsed this into an interactive web portal, which runs in our hospital and supervises a folder and it’s minutes, collects the data that we generate in the flow cytometry diagnostic setting, and sends that anonymized and encrypted to a supercomputer, and finally runs the algorithm and gives us the data through an interactive application on how likely is that sample to be pathogenic. So, how likely is that to be lymphoma? We are already planning next developments in integrating generative AI, so we can not only predict whether it’s tumor or not, but why, and which is the phenotype of these cells, and which is likely the diagnosis. But that’s probably for the next ASH.

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