JMIR Publications

ABSTRACT

Background:

Heparin is one of the most commonly used medication for patients during their ICU stay. In clinical practice, the use of a weight-based heparin dosing nomogram has been the standard practice for treatment of thrombosis. Recently, machine learning techniques have dramatically improved the ability of computers to provide clinical decision supports, and it raised the possibility of automatically generating computer-aided heparin dosing recommendations.

Objective:

The objective of this study is to recommend optimal heparin dosing strategies in Intensive Care Unit (ICU) using machine learning methods. Heparin is a common drug in ICU which is frequently misdosed since most medication guides taking only patient weight into account.

Methods:

Retrospective data from multi-parameter intelligent monitoring in intensive care III (MIMIC-III) were used as analysis dataset. 2789 patients who met the inclusion criteria from MIMIC-III database were included and grouped by administrations. Several clinical features which may affect heparin treatment effects were selected. 5 candidate machine learning models were compared in two patient groups to evaluate the classification performance for predicting 3 patient states of sub-therapeutic, normal-therapeutic and supra-therapeutic. The model results were evaluated by classification precision, recall, f-score and accuracy.

Results:

In our 3-class classification task, the shallow neural network algorithm performed best. In two patient groups, the shallow neural network model has achieved the model accuracy to 88.00% and 86.25% for patients accepted intravenous push and intravenous drip, respectively, which are both higher than performances of other models. The model results outperformed general heparin treatments by manual analysis and validating.

Conclusions:

The study is to find the most appropriate model to optimize the heparin dosing by comparing multiple machine learning models. We have separated the route of administration to further illustrate the model capability within different patient groups. Our study demonstrates the feasibility of recommending heparin dosing by data driven methods, thus to further improve heparin dosing and patient safety. Multi-center ICU data will be used in next step to further evaluate and improve the model results.