AI-powered precision: Unmasking hidden treatment optimizations for lower-risk myelodysplastic syndrome patients based on real-world, multi center EHR data.

Abstract

Abstract Introduction: Patients with lower-risk myelodysplastic syndrome (LR-MDS) benefit from treatment of clinically significant anemia with erythropoiesis-stimulating agents (ESAs) or erythroid maturation agents (EMAs). However, in real-life practice, the timing of initiation, discontinuation, and switching between therapies is often suboptimal. This study aimed to develop and implement an artificial intelligence (AI) algorithm, using data from electronic health records (EHRs), to identify patients within hospital databases who either remain untreated despite meeting criteria for ESA treatment or have had a suboptimal response to ESAs and may benefit from switching to EMA. The project was conducted between January 2024 and December 2024 and was supported by Bristol Myers Squibb Poland. Methods: We used the Saventic Med platform to design and test a multi-step, rule-based algorithm that integrates expert-driven logic with natural language processing (NLP) for extracting both descriptive and laboratory data. This project utilized anonymized EHRs and was approved by the Rzeszow University Ethical Commission (ethical assessment No. 2024/05/025). The algorithm was structured to identify and classify patients into two populations: (1) untreated LR-MDS patients potentially eligible for ESA with no documented ESA treatment in their EHRs, and (2) patients with an inadequate ESA response who may benefit from a treatment change. Features were extracted from EHR data using a hybrid Named Entity Recognition algorithm that combines expert knowledge with a BERT-based (Bidirectional Encoder Representations from Transformers) context recognition machine learning model. Subsequently, the extracted features were automatically analyzed using a highly explainable expert rule-based model. At each step, patient selection was guided by expert knowledge and current clinical guidelines. Results: The algorithm was implemented in 3 Polish hospitals. Two million anonymized patient EHRs were screened within 10 screenings. (1) The algorithm identified 63 patients potentially eligible for ESA treatment without information about such treatment. Based on EHR analysis and partial feedback from participating centers, 31 of 63 (49,2%) patients initiated ESA treatment during the project period. Out of them, 3 (4,8%) patients initiated treatment based solely on recommendations provided by the algorithm. The remaining 28 (44,4%) patients had initiated therapy concurrently or previously but this information was not present in EHRs. The follow-up for 32 (50,8%) patients has not yet been received. The concordance between the algorithm's recommendation and the physician's decision among patients with follow-up was 100%. (2) In addition, the algorithm flagged 34 transfusion-dependent patients receiving ESA therapy for suspected treatment failure. Following the report, ESA therapy was discontinued in 3 (8,8%) patients; all had ring sideroblasts (RS) >15% and subsequently initiated treatment with luspatercept. In 11 (32,4%) patients, treatment was modified either prior to or concurrently with the report. Twelve (35,3%) patients continued their previous treatment regimen despite the report. Four (11,8%) patients progressed to high-risk MDS or CMML, 3 (8,8%) patients were lost to follow-up and 1 (2,9%) patient died. The concordance between the algorithm's recommendation and the physician's decision among patients with follow-up was 41.2 %. Conclusions: The multi-stage AI algorithm used in this pilot project successfully identified LR-MDS patients who could benefit from modifications to their ongoing treatment. Despite the limited frequency of patient record updates (every 2–3 months), the algorithm was still able to detect individuals who had been overlooked within standard clinical workflows. Concordance between physicians and the algorithm shows that discontinuation of ESA therapy in transfusion-dependent patients may have the highest potential for optimization. The project has now been scaled up to eight medical centers, with EHR data updated on a monthly basis. The long-term objective is to enable real-time data integration. Generally, the project demonstrated the potential of EHR-based AI algorithms to assess guideline-concordant therapy in real-life.