DP019 | GENERATION OF PROGNOSTIC MODELS FOR ACUTE MYELOID LEUKEMIAS (AML) THROUGH INTEGRATED COMPUTATIONAL ANALYSIS OF ROUTINE CLINICAL DATA: A PILOT STUDY

Integrating hospital-generated clinical data has a high potential to optimize patient care, enhance prognostic accuracy, enable personalized treatment strategies for improved outcomes, and help in making data-driven health policies decisions. Nevertheless, several challenges remain, such as different data store formats interoperability, standardization issues, data privacy and security, data quality assurance. In the present study, we aimed at leveraging laboratory routine data at Fondazione IRCCS Ca’ Granda Ospedale Maggiore Policlinico, Milan, Italy, from 420 AML patients, diagnosed from 2011 to 2022 (male/female: 241/179, median age 65 yrs). All the data generated by routine laboratories were preliminarily categorized in the institutional data warehouse. Afterwards, all data, strictly anonymized, were managed and analyzed using KNIME Analytics Platform (Zurich, Switzerland). Demographic, outcome and laboratory variables including hematologic, coagulation and biochemistry parameters were analyzed by univariate and multivariate Cox regressions to identify prognostically relevant factors associated with overall survival (OS). For each significant variable, based on hazard risk, a risk score was assigned: age (>60 yrs, HR 5.3, score 1.5; >65 years, HR 2.6, score 1.5; >75 years, HR 2.0, score 1.0), direct bilirubin (HR 2.1, score 1.5), mean corpuscular hemoglobin concentration (HR 0.071, score -1.0), D-Dimer (HR 0.6, score -1.0), mean platelet volume (HR 1.6, score 1.0) and platelet count (HR 0.5, score -1.0). A prognostic model was generated that stratified AML patients into 3 OS risk groups: low (<0.5 points), medium (from <0.5 to <1.5 points), and high-risk groups (>1.5 points)(p<0.01) (Fig.1). This novel approach represents a “proof of concept” study that highlights the potential for the computational integration of clinical data, through their categorization in a data warehouse, allowing powerful and extremely efficient data extraction in terms of time and flexibility, as demonstrated by our experience, in which a strong prognostic model has been generated integrating a simple set of routine laboratory variables. Our future research will focus on validating and refining this prognostic model, in particular including molecular and cytogenetic variables, and to extend the use of this approach to different diagnoses, therapies, variables and clinical endpoint, as DFS, PFS, hospitalization duration, response to therapy.