Frailty as a predictor for outcome of Non-Invasive Respiratory Support in Acute Respiratory Failure: an innovative study supported by Machine Learning algorithms

Introduction Non-Invasive Respiratory Support (NIRS) is a crucial treatment option for Acute Respiratory Failure (ARF), though NIRS failure may lead to delayed advanced invasive support and therefore increased mortality. Frailty is a multidimensional syndrome due to depletion in physiologic reserves resulting in increased susceptibility to adverse outcomes and has been studied in the past in its relationship with ARF and NIRS outcome. Machine Learning (ML) is a field of Artificial Intelligence that provides output, such as failure probability, given a certain amount of input data. Materials and Methods This study was a multicentre retrospective analysis of data on 341 ARF patients treated with NIRS in Italy and in the UK. Primary aim was to investigate the predictive value of frailty in addition to conventional clinical data through the use of ML, exploratory analysis was to study its performance in specific subgroup, particularly the HFNC-only one. NIRS failure was defined as treatment escalation, death within seven days, or physiological failure. Frailty was assessed using the Clinical Frailty Scale (CFS) retrospectively using pre-admission functional information. Two ML models (TabPFN and XGBoost) were used, providing age and clinical data (respiratory rate and FiO2) within 6 hours before NIRS initiation and 1-2 hours after, and through feature ablation the impact of adding pre-admission CFS was quantified. A logistic regression model using CFS alone was also used. Performance was quantified through repeated 10-fold cross-validation; Area Under the Curve (AUC), Sensitivity, Specificity, Accuracy, Balanced Accuracy and Brier Score were calculated and compared by the corrected t-test. SHapley Additive exPlanations (SHAP) analysis was used to show the weighted contribution of each input variable to the output. Results Frailty was highly prevalent in NIRS failure without statistical significance (AUC=0.59±0.07, p=0.065), but feature ablation proved that CFS was significantly influent in ML predictive performance increasing AUC from 0.77±0.08 to 0.82±0.07 in TabPFN (p=0.001) and from 0.77 ± 0.08 to 0.81 ± 0.06 for XGBoost (p=0.001). SHAP analysis showed CFS to be the second most important determinant in both ML models. Exploratory analysis in HFNC-only subgroup showed an even more enhanced predictive performance through CFS inclusion (AUC increased from 0.889 ± 0.08 to 0.946 ± 0.035, p=0.009). Conclusions CFS appears to yield relevant influence on ML predictive performance, showing complementary properties to age and conventional clinical data. Combination of all these data may cover all stages of development of ARF, from baseline vulnerability to dynamic physiological response. Furtherly enhanced predictive performance in HFNC-only subgroup may support the hypothesis of interpreting HFNC as a NIRS form that is more dependent on residual functional capacity which is lower in frail patients. Although these hypothesis-generating findings require prospective validation, they support more widespread recording of CFS in acute respiratory care.