JMIR Publications

ABSTRACT

Background:

Although many machine learning models have been proposed to help clinicians predict outcomes after pediatric heart surgery, most models do not take full advantage of the detailed cardiac assessment information from echo report, and the interpretability and operability of the models are relatively poor.

Objective:

To address these challenges, a patient similarity network (PSN) of congenital heart disease (CHD) was proposed to provide a general prediction of outcomes after congenital heart surgery.

Methods:

Data from 4774 CHD surgeries were collected. A total of 66 indicators and all diagnoses will be extracted from each echocardiographic report using natural language processing technology. Combining several basic clinical and surgical information sources, the distances between each patient were measured by a series of calculation formulas. Inspired by structural mapping theory, the fusion of distances between different dimensions can be modulated by clinical experts. A user-operable PSN of CHD called CHDmap was proposed and developed.

Results:

Using 256 CHD cases, CHDmap was evaluated on two different types of postoperative prognostic prediction tasks: a binary classifications task to predict postoperative complication; a multiple classification task to predict mechanical ventilation duration. A simple poll of the k most similar patients provided by the PSN can achieve better prediction results than the average performance of three clinicians. Constructing logistic regression models for prediction using similar patients obtained from PSN can further improve the performance of the two tasks (best AUC=0.810 and 0.926, respectively).

Conclusions:

Without individual optimization, CHDmap demonstrates competitive performance compared to clinical experts. Furthermore, the advantage of CHDmap is that it can be combined with the clinician’s advanced cognition to make collaborative decisions. The generality, interpretability, and manipulability of the CHDmap made it more acceptable to clinicians.