Hybrid Health Recommender Systems for Antibiotic Prescription in Sepsis Patients: Statistical Method Development and Evaluation
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Patient survival in sepsis critically depends on immediate and appropriate empiric antibiotic therapy. Selecting the optimal therapy is challenging in practice due to often incomplete information and time constraints, while minimizing broad-spectrum antibiotic use is essential to prevent resistance development. Clinical decision support systems (CDSS) can assist physicians in deciding on an empiric therapy.
Methods
We explore health recommender systems (HRS) as a methodological basis for such CDSS with the aim to predict the therapy response of newly diagnosed sepsis patients to all eligible therapies. Using retrospective clinical and microbiological data from a German university hospital, we introduce and evaluate four HRS of varying complexity, input information and applicability. Each system combines collaborative and demographic filtering in a hybrid manner, employing both memory- and model-based filtering techniques. The HRS make use of both patient similarity (defined by core data and laboratory values) and therapy similarity (defined by pathogen coverage).
Results
The best results regarding precision, and F1 score are achieved by the least complex HRS (using memory-based filtering with patient similarity) and the most complex HRS (using model-based filtering and integrating patient core data and laboratory values). However, performance varies significantly across individual therapies. Memory-based methods are particularly suitable for new patients without prior therapy information if data is balanced, whereas model-based approaches perform better when therapy response is already partially available.
Conclusions
We present a methodological framework for hybrid HRS that aligns both memory-and model-based approaches with their respective optimal clinical scenarios, while underscoring the critical necessity for high-quality data to handle patient heterogeneity. While application in clinical practice is not yet feasible based on current data, our proposed approaches advance sepsis therapy research and offer a foundation for future studies.
