Machine learning probability calibration for high-risk clinical decision-making

SummaryMaking decisions in the context of risk is an integral part of psychiatric work. Despite this, decision-making skills are rarely systematically taught and the processes behind decisions are rarely made explicit. This article attempts to apply contemporary evidence from cognitive and social psychology to common dilemmas faced by psychiatrists when assessing and managing risk. It argues that clinical decision-making should acknowledge both the value and limitations of intuitive approaches in dealing with complex dilemmas. After discussing the various ways in which clinical decision-making is commonly derailed, the article outlines a framework that accommodates both rational and intuitive modes of thinking, with the aim of optimising decision-making in high-risk situations.

Download Full-text

579. Machine-Learning Based Models for Prediction of Recurrence-free Catheter Retention After ALT Treatment of CLABSI in a Pediatric Population

Open Forum Infectious Diseases ◽

10.1093/ofid/ofz360.648 ◽

2019 ◽

Vol 6 (Supplement_2) ◽

pp. S273-S274

Author(s):

Lorne W Walker ◽

Andrew J Nowalk ◽

Shyam Visweswaran

Keyword(s):

Machine Learning ◽

Decision Making ◽

Blood Culture ◽

Clinical Decision Making ◽

Pediatric Population ◽

Recurrent Infection ◽

Clinical Decision ◽

Definitive Treatment ◽

Support Vector ◽

The Future

Abstract Background Deciding whether to attempt salvage of an infected central venous catheter (CVC) can be challenging. While line removal is the definitive treatment for central-line associated bloodstream infection (CLABSI), salvage may be attempted with systemic antibiotics and antibiotic lock therapy (ALT). Weighing risk and benefit of CVC salvage is limited by uncertainty in the future viability of salvaged CVCs. If a CVC is likely to require subsequent removal (e.g., due to recurrent infection) salvage may not be beneficial, whereas discarding a viable CVC is also not desirable. Here we describe a machine learning approach to predicting outcomes in CVC salvage. Methods Episodes of pediatric CLABSI cleared with ALT were identified by retrospective record review between January 1, 2008 and December 31, 2018 and were defined by a single positive central blood culture of a known pathogen or two matching cultures of a possible contaminant. Clearance was defined as 48-hours of negative cultures and relapse was defined as a matching positive blood culture after clearance. Predictive models [logistic regression (LR), random forest (RF), support vector machine (SVM) and an ensemble combining the three] were used to predict recurrence-free CVC retention (RFCR) at various time points using a training and test set approach. Results Overall, 712 instances CLABSI cleared with ALT were identified. Demographic and microbiological data are summarized in Tables 1 and 2. Few (8%) instances recurred in the first 28 days. 58% recurred at any time within the study period. Rates of RFCR were 75%, 43%, 22% and 10% at 28, 91, 182 and 365 days. Machine learning (ML) models varied in their ability to predict RFCR (Table 3). RF models performed best overall, although no model performed well at 91 days. Conclusion ML models provide an opportunity to augment clinical decision making by learning patterns from data. In this case, estimating the likelihood of useful line retention in the future could help guide informed decisions on salvage vs. removal of infected CVCs. Limitations include the heterogeneity of clinical data and the use of an outcome capturing both clinical decision making (line removal) and infection recurrence. With further model development and prospective validation, practical machine learning models may prove useful to clinicians. Disclosures All authors: No reported disclosures.

Download Full-text