Page 54 - Read Online

Page 203 Jabbari et al. Art Int Surg. 2025;5:200-9 https://dx.doi.org/10.20517/ais.2024.77

Affordable and convenient PAD screening may offer significant benefits in various clinical settings. Kim
et al. performed one of the first proof-of-concept studies using deep CNN to detect and assess the severity
[30]
of PAD based on brachial and arterial pulse waveforms . Their work showed that DL may diagnose PAD
more accurately compared to current ankle-brachial index techniques. Allen et al. further demonstrated the
value of DL-based approaches in PAD screening . Their team used DL-based photoplethysmography
[31]
(DLPPG) classification to achieve high diagnostic performance with toe PPG signals. Within this portable
and inexpensive model, data are transmitted to servers where DL algorithms facilitate accelerated and
accurate diagnoses of PAD.

Additionally, timely detection of diabetic foot ulcers is critical in preventing LLA. Several reviews have
reported on the application of AI in the diagnosis and treatment of diabetic foot [32,33] . A recent proof-of-
concept study by Cassidy et al. demonstrated accurate diabetic foot ulcer detection with an AI system on
smartphones . A total of 203 photographs were taken by smartphone, analyzed by the AI system, and
[34]
compared against expert clinical judgment. The predictions and decisions made by the AI system displayed
high sensitivity (0.92) and specificity (0.86).

AI in patient management and clinical decision making
The application of AI in clinical decision making may revolutionize surgical practice through novel patient-
[35]
centered approaches. Chung et al. used ML to generate an accurate risk prediction model for CLTI . Their
multicenter, nested study included clinical trial data from 1,238 patients undergoing infrainguinal vein
bypass for the treatment of ischemic rest pain or ischemic tissue loss. Supervised topic model cluster
analysis was able to differentiate three distinct clusters of patients within the nested cohort, each designated
as a specific stage within CLTI. Cluster analysis revealed 1-year CLTI-free survival rates of 82.3% for stage 1,
61.1% for stage 2, and 53.4% for stage 3. Stratification by stage revealed major limb amputation rates of 4.2%
for stage 1, 10.8% for stage 2, and 18.4% for stage 3. Among those without a major amputation, the rate of
CLTI recurrence was directly related to increasing stage number. Similarly, Oei et al. developed ML
algorithms to predict the risk of LLA in 2,559 patients with diabetic foot ulcers . Their model performed
[36]
well in the prediction of major [area under the receiver operating characteristic curve (AUROC): 0.820],
minor (AUROC: 0.637), and any (AUROC: 0.756) LLA events. They further determined total white cell
count, comorbidity score, and red blood cell count as key factors associated with the risk of major
amputation. The above studies depict emerging methods for risk stratification and outcome prediction,
highlighting the power of AI applications in surgical decision making.

The management of the mangled extremity represents yet another complex decision-making scenario. The
decision for amputation or limb salvage will likely be innovated by AI models and replace traditional
scoring systems . Perkins et al. developed a Bayesian network (BN) prediction model using a supervised
[37]
ML approach to estimate the outcome of limb revascularization, a metric often critical to attempting limb
salvage versus amputation . The prediction model sourced information from domain knowledge,
[38]
published data, and US Department of Defense Data. Their model accurately predicted failed
revascularization (AUROC: 0.95), with maintained performance on external validation (AUROC: 0.97). The
BN prognostic model outperformed the traditional mangled extremity severity score in predicting the need
for amputation [AUROC: 0.95 (0.92-0.98) vs. 0.74 (0.67-0.80); P < 0.0001].

Following the decision to perform a procedure, surgeons are often faced with postoperative patient
complications. In general, the perioperative period serves as the source of initial exposure for many patients
with chronic opioid use [39,40] . Using a ML approach, Gabriel et al. developed predictive models for persistent
opioid use following lower extremity joint arthroplasty . They demonstrated that ensemble learning can
[41]