Smart, automated junctional tourniquets leveraging AI-driven ultrasound guidance

Abstract

Tourniquets are commonly used devices for hemorrhage control; however, their effectiveness is reduced in anatomical junctions such as the neck and inguinal region. Junctional tourniquets specifically require precise placement to be effective. This precision can be enabled with ultrasound technology to help locate and occlude the major vessels in the junctional regions properly. However, interpretation of ultrasound requires highly skilled personnel, who may not necessarily be available in emergency situations. To overcome this hurdle, we have developed two ultrasound-enabled, AI-driven junctional tourniquet prototypes. AI models can aid in guiding the end-user to the correct location and determine occlusion during and after pressure application. Proof-of-concept functionality of the developed prototypes integrated with AI models was successfully tested in a durable, ultrasound-compatible femoral tissue phantom and compared against commercially available tourniquet devices. Overall, time to occlusion was comparable between the tourniquet prototype designs and traditional junctional tourniquets, while each AI model achieved high performance metrics for this application. As such, the combination of AI and ultrasound can prove to be a viable solution to prevent further hemorrhaging at the point of injury.

Fig. 1

Summary of testing performance for each prototype and commercial tourniquet for (A) subclavian, (B) aorta, and (C) femoral placements. Times are shown as average for the assembly, securement, and occlusion period in stacked bar plots, with error bars denoting the standard deviation for each (n = 3 replicates) run that achieved proper placement. (D) Validation of flow rate changes for all sites during baseline, occlusion, and release of pressure. Statistically significant differences between overall tourniquet times were determined by ordinary one-way ANOVA, post-hoc Holm-Šídák test (* for p-values < 0.05; ** for p-values < 0.01). Flow rates are shown in arbitrary units with the height of the box representing the first and third quartile and error bars denoting range of values.

Fig. 2

Representative images of FRejT (First row) and BaTS (Second row) placement at the Subclavian (First column), Aortic (Second column), and Femoral (Third column) sites in the commercial tissue phantom.

Fig. 3

Summary of guidance AI model performance metrics. Anatomical guidance object detection model performance metrics on blind test subjects for (A) Precision, (B) Recall, (C) F1 Scores, and (D) IOU. Artery, Vein, and Bone are shown separately alongside overall performance metric scores. Results are shown as box and whisker plots, with error bars indicating minimum and maximum values while the length of the box indicates interquartile range and median value is denoted by the line (N = 4 US video captures).

Fig. 4

Summary of occlusion AI model performance metrics. (A) Comparison of performance accuracy for different occlusion AI classification model configurations. Results are averaged across data capture experiment starting at baseline, progressive pressure, and reaching full occlusion, as indicated by the bottom “Flow Rate” figure panel. Average accuracies are shown for the multi-class, three-class, two-class, and qualitative image classification models. (B) Summary of correct and incorrect predictions across all test images (N = 754 frames) for each occlusion AI model. Statistical differences between occlusion AI models were determined by McNemar test with statistically significant (**** denotes p-value < 0.0001) and not significant results shown (ns).

Fig. 5

Timing to reach occlusion. (A) Representative ultrasound images from JUGO phantom categorized as “baseline” or “occlusion” based on the qualitative split criteria. Image brightness and contrast were adjusted for visualization purposes. (B) Summary of times for detecting occlusion by the AI classification model and the additional time to reach objective occlusion, defined as 90% flow reduction. Four individual replicates are shown to highlight the performance consistency.

Fig. 6

Proof-of-concept testing results for real-time prototypes. (A) Comparison of performance between normal, frame pooling, and single vessel object detection configurations across n = 10 replicates real-time runs for BaTS and FRejT. Differences between guidance timings were not significant as determined by Kruskal-Wallis, post hoc Dunn’s test. (B) IOU scores for each captured frame for bone, vein, and artery features. Significance differences between groups (* p-value < 0.05; *** p-value < 0.001; **** p-value < 0.0001) were determined by Kruskal-Wallis test, post hoc Dunn’s test. Overview of (C) time to occlusion and (D) occlusion accuracies for BaTS and FRejT. Differences between groups were not significant as determined by unpaired T-Test.

Fig. 7

Design of the Frame Reinforced Junctional Tourniquet (FRejT). A computer-aided design overview of the FRejT prototype is shown along with the labels for the (A) aluminum base plate, (B) vertical t-slot extrusion, (C) 3D-printed jig for attachment to (D) horizontal t-slot extrusions, (E – G) main housing of the US probe comprised of (E) outer securing layer, (F) inner moving layer, and (G) probe holder.

Fig. 8

Design of the Base and Tightening Straps Junctional Tourniquet Prototype. A computer-aided design overview of the BaTS prototype is shown along with the labels for the (A) inner mold, (B) external chassis, (C) aluminum collar, (D) motor anchoring cap, (E) straps, (F) ultrasound probe, and (G) mounting points for handle or electronic attachment. The different views of the prototype show the extension of the US probe by linear actuation during occlusion and collar rotation functionality.