Relating process and outcome metrics for meaningful and interpretable cannulation skill assessment: A machine learning paradigm.

• Based on data collected on our cannulation simulator, the relationship between process metrics and outcome metrics for this task is explored. • The above modeling is undertaken assuming that most of the process metrics and the outcome metric are on a continuum (versus discrete values), allowing for fine-grained skill assessment. • From the machine learning-based models, we seek to identify the most salient process metrics that impact outcomes for cannulation skill improvement. This knowledge could be used to develop personalized training strategies for skill progression over time Background and Objectives : The quality of healthcare delivery depends directly on the skills of clinicians. For patients on hemodialysis, medical errors or injuries caused during cannulation can lead to adverse outcomes, including potential death. To promote objective skill assessment and effective training, we present a machine learning approach, which utilizes a highly-sensorized cannulation simulator and a set of objective process and outcome metrics. Methods : In this study, 52 clinicians were recruited to perform a set of pre-defined cannulation tasks on the simulator. Based on data collected by sensors during their task performance, the feature space was then constructed based on force, motion, and infrared sensor data. Following this, three machine learning models– support vector machine (SVM), support vector regression (SVR), and elastic net (EN)– were constructed to relate the feature space to objective outcome metrics. Our models utilize classification based on the conventional skill classification labels as well as a new method that represents skill on a continuum. Results : With less than 5% of trials misplaced by two classes, the SVM model was effective in predicting skill based on the feature space. In addition, the SVR model effectively places both skill and outcome on a fine-grained continuum (versus discrete divisions) that is representative of reality. As importantly, the elastic net model enabled the identification of a set of process metrics that highly impact outcomes of the cannulation task, including smoothness of motion, needle angles, and pinch forces. Conclusions : The proposed cannulation simulator, paired with machine learning assessment, demonstrates definite advantages over current cannulation training practices. The methods presented here can be adopted to drastically increase the effectiveness of skill assessment and training, thereby potentially improving clinical outcomes of hemodialysis treatment. [ABSTRACT FROM AUTHOR]