JMIR Publications

ABSTRACT

Background:

Point-of-care ultrasonography to assess the diaphragm has seen rapid growth in practice in critically ill patients of invasive mechanical ventilation. Previous research developed an automatic measurement of diaphragmatic excursion and velocity based on two-dimensional speckle-tracking technology

Objective:

The study aimed to develop artificial intelligence-multimodal learning to improve performance in predicting weaning failure and guiding the accurate weaning process.

Methods:

Artificial intelligence-multimodal learning based on clinical characteristics, laboratory parameters, and diaphragm ultrasonic videos was established. Clinical indicators, diaphragm ultrasound video, and diaphragmatic indicators (using automatic speckle-tracking measurement) were collected to construct clinical/diaphragmatic dataset and diaphragm ultrasound dataset. Four experiments were conducted to predict weaning failure in an ablation setting, testing each component of the Co-learning model with corresponding data. The prediction performance was compared in a multi-modal setting with three single-modal settings (diaphragmatic excursion, clinical/diaphragmatic indicators, and diaphragm ultrasound video) by classification accuracy, ROC curves, Precision-recall Curve, and Calibration Curve.

Results:

The accuracy improved when predicted through diaphragm ultrasound video data using ViViT (Accuracy=0.8095, AUC=0.852), clinical/ultrasound indicators (Accuracy=0.7381, AUC=0.746), and the multi-modal Co-learning (Accuracy= 0.8331, AUC=0.894). The proposed Co-learning model achieved the highest score (AP=0.91) among the four experiments. The calibration curve demonstrated that the proposed Co-learning model was well calibrated as the curve was closest to the perfectly calibrated line.

Conclusions:

Using both ultrasound and clinical data for Co-learning improved the accuracy of the weaning outcome prediction. Multi-modal learning based on automatic measurement of POCUS and automated collection of objective clinical indicators greatly enhanced the practical operability and user-friendliness of the system, which was conducive to clinical promotion.