Red blood cell transfusion, patient outcomes and anaemia in elective orthopaedic surgery through the lens of supervised learning and mediation analysis

Introduction: An urge in the patient blood management and transfusion medicine landscapes to study patient outcomes is prevalent in pursuit of relieving transfusion dependency and enhancing patient-centredness. Clinical researchers prefer simple models and ease in their interpretability. The incentive by the Sanquin organization is to conduct mediation analysis – novel in this field of medicine. Literature offers studies with blood transfusion treated in terms of the product use, not yet through the lens of patient outcomes. A raw, patient-level dataset from the ‘TOMaat' study (a double-randomized multi-centre control trial) from the elective orthopaedic surgery featuring 533 variables was available for this research. Methodology: The research examines the role of red blood cell (RBC) transfusion up to Day 14 relative to postoperative complications up to Day 14. Pre-operative anaemia is the exposure component in the mediation model. A blend of prediction and inference tools were utilized in supervised machine learning model development and mediation analysis on a sample size of 2426 patients. Partial dependence plots, odds ratios and coefficients yielded effect estimates from non-parametric (random forest (RF)) and parametric models (logistic regression (LREG) and lasso). The raw dataset was subject to thorough variable selection to reduce the number of input variables from 533 to 41 and 32. Respectively, this applies to the models with post-operative complication up to Day 14 (Case COM) and RBC transfusion up to Day 14 (Case RBC) being the target dependent variables. Lasso led to a further reduction of input variables to 11-12 and 8 (COM and RBC, respectively). Due to excessive data missingness (34% in COM) and a free text field format (RBC) of event dates, massive data cleaning efforts led to establishing pessimistic and optimistic scenarios (COMPES, COMOPT) to sequence the RBC and COM events in time. Results: All 12 supervised learning models display moderate performance in terms of the AUC (0.63-0.71) with no significant difference between the RF, tuned RF, LREG and lasso models (built per each Case RBC, COMPES, and COMOPT). Strong confounding variables were consolidated from the inference insights and thoroughly validated with the clinical expert leading to 10 strong confounders for the mediation model. RBC transfusion is a statistically significant predictor for COMPES based on LREG and lasso; however, opposing results are found for COMOPT. Different results may be observed when examining pre-operative anaemia in silo using descriptive statistics (p<0.001 for COMPES and COMOPT, chi-squared test) versus in the presence of other covariates resulting in low variable importance based on the supervised learning models. Extreme implications due to data missingness were visible in mediation analysis since opposing findings were observed for these two scenarios. RBC transfusion mediates the relationship between pre-operative anaemia and post-operative complications in the pessimistic scenario (ACME of 0.0445, 95% CI of [0.0268; 0.0700]) whereas there is no significant mediation in the optimistic scenario.