Skip to main content
IMS Events Main banner
Back

    Myeloma Genomics and Microenvironment and immune profiling

    Artificial Intelligence system to predict t(11,14) status in multiple myeloma patients by leveraging the correlation of plasma cell morphology and genetics

    P-199: Artificial Intelligence system to predict t(11,14) status in multiple myeloma patients by leveraging the correlation of plasma cell morphology and genetics

    Thursday, September 26, 2024
    Introduction: The diagnosis of multiple myeloma (MM) is based on the cytomorphological analysis of bone marrow aspirate (BMA), along with cytometry and genetic analysis based on underlying cytogenetic alterations, which have both prognostic and therapeutic implications. An example is the translocation (t)(11;14), which, although it does not have a prognostic impact, serves as a therapeutic target by allowing targeted treatment with venetoclax in these patients. The objective of our study is to develop an artificial intelligence (AI) algorithm to automatically identify plasma cells (PC) with t(11;14) from morphological features of BMA samples analyzed using automated optical microscopy.

    Methods: We digitized 40 fields (100x) from each patient, including healthy controls, controls with MM without t(11;14), and patients with confirmed MM with t(11;14), using a device manufactured with a 3D printer to attach a mobile application to an optical microscope. For each BMA, 40 fields (100x) were digitized and the average of PC analyzed was 174. An AI algorithm was developed based on a database of more than 400,000 triple manually labeled cells, capable of automatically identifying nucleated cells from BMA images and classifying them into PC and non-PC, with an area under the curve (ROC-AUC) of 98%. Subsequently, an AI algorithm was developed to learn morphological patterns of PC associated with t(11;14) and predict its presence. The AI algorithm corresponds to a Multiple Instance Learning (MIL) architecture, which in turn uses features of each cell extracted by a foundational model trained through self-supervised learning on a database of more than 100K BMA cells. This MIL algorithm for predicting t(11;14) was trained using a 3-fold cross-validation technique, thus ensuring a division at the patient level between training and validation sets.

    Results: A total of 881 digitized cases were analyzed, 841 negative [677 healthy controls and 164 MM without t(11;14)] and 40 MM positive for t(11;14). Morphological analysis of plasma cells using the AI algorithm proved effective for detecting PCs with t(11;14) with an area under the ROC curve (ROC-AUC) of 0.88. The algorithm can predict the presence of t(11;14) with a sensitivity of 0.83 and a specificity of 0.74. The false positive rate was 0.26, while the false negative rate was 0.17.

    Conclusions: Our results demonstrate the potential of applying AI models in the morphological analysis of PC in BMA for detecting genetic alterations, confirming the morphology-genetics correlation in PC of MM cases with t(11;14). Our algorithm sensitively identifies the presence of morphological features suggestive of t(11;14); however, it shows more limited specificity, which could likely be resolved by increasing the sample size, refining the model architecture, or exploring alternative models that include other predictive variables.