P1213: ARTIFICIAL INTELLIGENCE TO PREDICT MEDICAL DIAGNOSIS FROM FLOW CYTOMETRIC RAW DATA IN MATURE B-CELL AND T-CELL NEOPLASMS, AML, ALL, MDS AND MULTIPLE MYELOMA
Martha‐Lena Müller, Sven Maschek, Elena Fortina, Marcelo Cunha, Simon Pagezy, Adriane Koppelle, Claudia Haferlach, Manja Meggendorfer, Torsten Haferlach, Wolfgang Kern · HemaSphere · 2023
Topic: 18. Indolent and mantle-cell non-Hodgkin lymphoma - Clinical Background: Artificial intelligence (AI) is becoming a key component in diagnostic workup solutions worldwide. Regarding our vast database of flow cytometry data of hematologic neoplasms, and the ever-growing amount of newly incoming patient samples, we developed and start to live-test AI-powered workflows to support high-throughput routine diagnostic procedures Aims: AI-based support of flow cytometric data analysis and interpretation of hematologic neoplasms Methods: We developed six AI models to decipher raw flow cytometry data to classify mature B-cell and T-cell neoplasms (B-NHL, T-NHL), acute myeloid leukemia (AML), acute lymphoblastic leukemia (ALL), myelodysplastic neoplasms (MDS) and multiple myeloma (MM). All models were trained on uniformly processed samples analyzed by human experts during standardized routine diagnostics (Navios/ Cytoflex cytometers yielding.lmd. or.fcs files, Kaluza software, Beckman Coulter, Miami, FL). For our models 21,769 cases were used. The models are built with XGBoost, use raw data without image recognition steps and utilize expert-informed features as input. Recall (R), precision (P) and prediction probabilities (PP) were recorded. Meanwhile, a user interface has been built and we have started to live-test newly incoming patient samples with suspected B-NHL on our internal AI platform. The models return a predicted diagnosis and its PP, and the decision process is displayed (Fig. 1). The predicted diagnosis is compared with the human expert diagnosis and this information is fed back to the model for continuous improvement. Results: For B-NHL, our most frequent flow cytometry analysis, the model was trained on 10,750 cases. It distinguishes 1) CLL/CLL-like and mantle cell lymphoma (MCL, n=3,990), 2) follicular lymphoma (FL, n=126), 3) hairy cell leukaemia (HCL) and splenic B-cell lymphoma/leukaemia with prominent nucleoli (SBLPN, n=702), 4) lymphoplasmacytic lymphoma (LPL) and marginal zone lymphoma (MZL, n=2,591) and 5) no lymphoma (n=3,342). Considering the top 75% cases (PP>0.87) the average R (aR) was 91% and the average P (aP) was 97% (Fig. 1a) on the dataset. During live testing, corresponding values for B-NHL were aR=99% and aP=87% (Fig. 1b; categories FL and HZL/SPLPN not shown: too small sample numbers). For the other AI models, considering the top 75% of cases, the performances were: AML (n=2,400) with categories 1) AML (n=1,706) and 2) no AML (n=694), aR=99%, aP=99%; ALL (n=220) with categories 1) Pro-B-ALL (n=25), 2) c-ALL (n=82), 3) cortical T-ALL (n=16), 4) non-cortical T-ALL (n=26) and 5) no ALL (n=71), aR=92% and aP=93%; MDS (n=3,234) with categories 1) MDS (n=1,804) and 2) no MDS (n=1,430), aR=94% and aP=94%. MM (n=4,754) with categories 1) MM (n=1,256), 2) consistent with MM (n=829), 3) consistent with MGUS (n=1,903), 4) no MM (n=766), aR=88% and aP=88%; T-NHL (n=411) with categories 1) T-/NK-cell neoplasia (n=123) and 2) no T-/NK-cell neoplasia (n=288), aR=93% and aP=94%. Summary/Conclusion: AI is applicable to analyze and interpret flow cytometric data allowing prediction of diagnostic results at high recall and precision. Application of AI solutions will enable us to shift particularly repetitive manual workflows away from humans who instead can focus more on using their expertise to secure more difficult and complex diagnoses. Further strengthening of the models, including more subtypes of hematological entities and rare event analysis in the MRD context will result in their broader applicability. These solutions will be made available to flow laboratories worldwide.Keywords: Artificial intelligence