Automated AI classification in clinical flow cytometry: Transforming B-cell lymphoma diagnostics

Abstract Introduction: Flow cytometry (FC) holds a pivotal role in hematological diagnostics. Although capable of turn-around time within hours, the manual processing of these complex data remains challenging and resource consuming. Integration of artificial intelligence (AI) algorithms for the analysis and interpretation of FC data is poised to revolutionize the field by unlocking previously unexplored capabilities. Aims: 1) Development of an AI classification model for interpretation of flow cytometric data and identification of B-cell lymphoma (B-NHL) cases as well as B-NHL subclassification with high accuracy, 2) reliable automated visualization, 3) routine implementation. Methods: An XGBoost-based classification model was trained and validated using FC data from 12,015 cases, consisting of 5,015 B-NHL and 7,000 no B-NHL cases. Lymphoma diagnoses included: monoclonal B-cell lymphocytosis/chronic lymphocytic leukemia [MBL/CLL]: 2,000, marginal zone lymphoma/lymphoplasmacytic lymphoma [MZL/LPL]: 1,095, mantle cell lymphoma [MCL]: 809, hairy cell leukemia [HCL]: 795 and follicular lymphoma [FL]: 316), respectively. The ground truth of FC set diagnoses was further underpinned by relevant genetic data. The cloud-based Cytobank data analysis platform (Beckman Coulter, Miami, FL) was used for standardized gating and transformation of raw FC files prior to analysis by the in-house AI classification model. Cytobank's AutoGating algorithm was trained on 50 manually gated cases, enabling single cell labeling and generation of labeled datasets. The prediction by the in-house AI model was deployed in two stages, i.e. binary classification (B-NHL vs. no B-NHL) and identifying B-NHL subtypes. In addition, an in-house automated tool was deployed to visualize cell populations in scatter plots corresponding to manual processing. Classification model was trained with a training cohort (n=9,612), its performance was verified by a validation cohort (n=2,403). Subsequently the model was prospectively challenged with new data from routine workflow (n=915), where lymphoma was part of differential diagnosis. Concordance between manual and automated process was assessed in a blinded manner. From the 915 cases only those with high prediction probability (>90%) were used for the model's evaluation in order to reflect future application, i.e. routine operations' demand to minimize additional manual processing. Results: In the validation cohort, a total of 1,802/2,403 (75%) cases showed prediction probability above 75% and were used for assessment of the model. Overall, very high accuracies of 99.3% and 98.7% were achieved for both B-NHL detection and subclassification. The prediction accuracies for the respective subtypes were: MBL/CLL: 98.6%, MZL/LPL: 96.4%, MCL: 98.1%, HCL: 99.2% and FL: 100%. As to prospective evaluation with new data, 672/915 (73%) cases were predicted with probability above the set threshold of 90%. Classification accuracies were very high also in this context and reached 98.7% for B-NHL detection and 97.0% for B-NHL subclassification. Again, accuracies were very high for MBL/CLL (98.1%), MZL/LPL (100%), MCL (100%), and FL (100%) identification, respectively, with sole exception being HCL with an accuracy of 86.9%. The automatically generated dot plots were rated by diagnostic experts as qualitatively appropriate and equivalent to manual data analysis in all 672 cases. Also, diagnostic experts made identical diagnoses in all 672 cases, respectively, when reviewing both automatically generated and manually prepared dot plots. In particular, review of automatically generated dot plots enabled them to set the correct diagnosis in all out of 16 cases wrongly predicted by the AI algorithm. Compared with manual processing, the automated workflow reduced hands-on-time by up to 75%. Conclusion: The prediction performance of the AI classification model to detect B-NHL in FC was outstanding both in the validation cohort and in the real-life routine data: Further, the automatically generated dot plots enabled a correct diagnosis by the diagnostic expert at all times. Both these components are applicable to routine operations. Our data strongly support the integration of our AI lymphoma classifier model into our routine workflow, which will dramatically reduce the hands-on-time and open the way for further applications that are time-consuming and analyses requiring high expertise, such as minimal residual disease detection or immunoprofiling.