Technologies for Systematic Procedure Generation of Enhanced Wound Care Devices Through Discriminative Intelligence

Abstract

Employing natural language processing (NLP) and deep learning, this study assesses the quality and impact of computerized order-based protocol assignment for magnetic resonance imaging (MRT) procedures (DL). NLP methods were used to handle orders from approximately 116,000 MRT exams, including 200 different sub-specialized protocols (the “Local” protocol class). Separate DL models for “Local” protocols, 93 American College of Radiology (“ACR”) protocols, and 48 “General” protocols were trained on 70% of the processed data. The DL Models were evaluated in two inference modes: “auto-protocoling (AP),” which produces the highest recommendation, and “clinical decision support (CDS),” which returns up to ten protocols for radiologist assessment. The difference between the normalized output score of the associated neural net for the top two recommendations was used to compute and examine the accuracy of each protocol recommendation. For the “General,” “ACR,” and “Local” protocol classes, the top predicted protocol in AP mode was correct for 82.8 percent, 73.8 percent, and 69.3 percent of the test cases, respectively. In CDS mode, accuracy levels above 96 percent were achieved for all protocol classes. However, the proposed models have a minor, positive budgetary impact on large-scale imaging networks at current validation performance levels. DL-based protocol automation is possible with more generic protocols and can be customized to route significant fractions of tests for auto-protocoling. Improved algorithm performance is necessary to produce a feasible exam auto-protocoling tool for sub-specialized imaging exams, according to economic assessments of the evaluated algorithms.

Ähnliche Arbeiten