Constructing a knowledge-driven and data-driven hybrid decision model for etiological diagnosis of Ventricular Tachycardia
Keywords: Ventricular Tachycardia, knowledge-driven, data-driven, machine learning, hybrid model, decision-making
TL;DR: A knowledge-driven and data-driven hybrid decision model
Abstract: Background: Each of the two clinical decision support models, data-driven and knowledge-driven, has its own unique strengths and challenges, and at the same time, the development of People-Centric Artificial Intelligent (PCAI) clinical decision support urgently needs to explore the effective integration strategies of the two driven models.
Objective: Constructing a trustworthy and highly accurate hybrid decision model incorporating knowledge-driven and data-driven model, and applying it to the field of healthcare.
Methods: We collected authoritative clinical practice guidelines, expert consensus and medical literature in the field of cardiovascular diseases as knowledge sources and retrospectively collected electronic medical record information of patients with ventricular tachycardia (VT) from Fu Wai Hospital as a dataset. The knowledge-driven model constructs a clinical pathway using a knowledge rule-based approach, and the data-driven model constructs a multi-classification machine learning model for etiological diagnosis of VT based on real-world data. The hybrid model's uses the clinical pathway as the basic framework, and the machine learning model is embedded as a custom operator into the decision node of the process. The comparison metrics of the three models are precision, recall and F1 score.
Results: In this study, we selected three clinical guidelines as the knowledge source for the knowledge-driven models, as well as collected 1,305 patient data as the dataset. A total of five machine learning models were constructed and the best model was XGBoost model (precision, recall, and F1 were 88.4%, 88.5%, and 88.4%, respectively. The hybrid model adopts the knowledge-driven thinking, embedding the machine learning model into the decision-making node of the two layers of classification, respectively. The precision, recall and F1-scores for the knowledge-driven model were 80.4%, 79.1% and 79.7%; for machine learning model were 88.4%, 88.5%, and 88.4%; for hybrid model were 90.4%, 90.2% and 90.3%.
Conclusion: The results show that the strategy of integrating knowledge-driven and data-driven clinical decision-making models is feasible. Compared to the pure knowledge-driven and data-driven models, the hybrid model demonstrated higher accuracy, and all the decision-making results of the model were based on evidence-based evidence, which was closer to the actual diagnostic thinking of clinicians. This new generation of PCAI systems for clinical decision-making needs to be applied to a wider range of medical fields and rigorously validated in the future.
Submission Number: 12
Loading