At OneGuild, we recognize that some of the most impactful innovations come from the front lines of clinical science. In this contribution, Dr. Jayer Chung — a vascular surgeon-scientist at Baylor College of Medicine — shares how artificial intelligence is shaping new approaches to one of diabetes' most serious complications: chronic limb threatening ischemia (CLTI). His article details current efforts to improve limb preservation through AI, data integration, and smarter risk prediction.

Artificial Intelligence in Chronic Limb Threatening Ischemia (CLTI): An Update

Written by Jayer Chung, MD, MSc - Division of Vascular Surgery and Endovascular Therapy, Michael E DeBakey Department of Surgery, Baylor College of Medicine, Houston TX

Address for correspondence:
Jayer Chung, MD, MSc, FACS
Associate Professor
Associate Chief of the Division of Vascular Surgery and Endovascular Therapy
Chief of Vascular Surgery at Baylor/St. Luke’s Medical Center
Michael E DeBakey Department of Surgery
Baylor College of Medicine
One Baylor Plaza, MS 390
Houston TX, 77030
Phone: (713)798-8840 - Fax: (713)798-7851
Email: Jayer.Chung@bcm.edu

Chronic limb threatening ischemia (CLTI) is the most severe clinical manifestation of peripheral arterial disease (PAD).1 CLTI occurs when there is inadequate blood supply to maintain a viable limb.1 CLTI is a rapidly progressing epidemic. The prevalence of CLTI is 1.34% of the United States population,2 which, corresponds to a population of 4.56 million people. This is larger than the population of the city of Los Angeles, the second largest metropolitan area in the United States. The prevalence has increased ten-fold over the last twenty years,2 and is likely to continue to increase due to the increased prevalence of atherosclerotic risk factors, such as age, diabetes, hyperlipidemia, and hypertension.

CLTI is very common amongst diabetic patients. There is a two-fold increase in PAD among diabetics.3 Diabetes and PAD impart a synergistic risk of major amputation. There up to a five-fold increase in the risk of major amputation when combining diabetes and PAD.3 CLTI is also a marker of overall poor health. Less than half of diabetics with CLTI will survive four years from presentation.1,3

The challenge when deciding how to manage CLTI is predicting who will benefit from procedures to try and save the limb. In this context, we need improved risk-stratification systems.4 These will allow us to compare outcomes between different centers as well as permit equitable comparisons of clinical trials and devices. In addition, the patient-centric benefit of improved risk stratification is improved numeric literacy to help reduce confusion on the part of the patient. The ultimate goal is create a tool that can improve decision-making for physicians, patients, and their caregivers and thereby optimize outcomes.4 Modern recommendations vary widely, due to poor, often conflicting, data to guide decisions.5

Prior prediction tools are inadequate to predict the risk of major amputation. The area under the curve (AUC; ie the amount of the variability in the study base that is described by the model) for prior models is modest, approximately 0.60.4 These models are limited because they fail to incorporate the severity of the wounds and infection well. Also, they fail to incorporate data regarding the vascular anatomy and perfusion metrics. The most modern risk stratification system is the Society for Vascular Surgery Wound Ischemia and foot Infection (WIfI) score.6 This has been externally validated in multiple centers to predict limb salvage. The WIfI score remains limited, however, as it does not incorporate patient co-morbidites, vascular anatomic, or perfusion data.4

The heterogeneity of treatments and patient presentations create a monstrous amount of data when trying to create accurate risk prediction models. Traditional mathematic techniques are not adequate to handle the amount of data in these models. Newer mathematical techniques have been developed and are loosely termed artificial intelligence (AI). AI is a term that encompasses machine learning, as well as deep learning.7 Herein, are a few examples that describe some the early applications of AI upon risk-stratification in CLTI.

The first is a study we performed at Baylor College of Medicine, where we compiled the data from 10 centers, resulting in 2878 limbs.8 We used k-means clustering, which is a form of machine-learning to see if we could organize WIfI better than the original WIfI classification. In the original WIfI, the data results in stages that overlap considerably with one another with respect to the actual 1-year rates of major amputation. In contrast, by using machine-learning, we could organize the WIfI stages better, and thereby improve the prediction of which patients would benefit from revascularization.

Another form of machine learning recently utilized is called topic cluster analysis.9 This utilized algorithms that associate words from charts or large data sets, with specific variables and outcomes. In so doing, we could determine which variables, when combined with WIfI, would best predict outcome. In fact, in our analysis, the topic-cluster analysis out-performed traditional logistic regression modeling, by including more variables and was therefore more reliable.

Moving on to deep learning, there are three applications that we are working on. The first is to capitalize on newer randomized controlled trial schemes. The one that we are utilizing is a Sequential Multiple Assignment Randomized Controlled Trial (SMART).10 This has been utilized successfully in the management of oncologic problems. The problem is that there are so many treatment sequences that a patient may take before successfully healing a wound, or conversely, tragically succumbing to amputation. In traditional randomized trials, this results in crossover, which greatly limits the statistical power of the trial. With a SMART, we have multiple randomization points built in, after which, we can utilize deep learning (forest plots and Recurrent Neural Networks; RNN) to determine the optimal sequence of treatment for a given patient presentation. We are currently working on a multi-center National Institutes of Health (NIH) submission to examine the best sequences of care in low-risk CLTI randomizing patients to wound care versus revascularization.

There are also ongoing efforts to incorporate anatomic data into the risk-stratification schemes. This requires deep learning, in the form of convoluted neural networks (CNN), recurrent neural networks (RNN) and visual transformation (ViT).7 The ultimate goal is to create a model that combines the patient’s data in the electronic medical record with the angiogram to develop a real-time virtual assistant, similar to Alexa, or other voice-activated systems.

Finally, we are similarly utilizing deep learning to enhance the management of depression in CLTI. Recent data from our center shows that severe depression afflicts up to one-third of CLTI patients. Yet, those that are treated remain low, at approximately twenty percent. Moreover, tragic events, such as suicide and homicide occur out of hospital, where we cannot detect them. We are therefore working to develop algorithms that identify the patient’s risk of suicide based upon clinical factors as well as data taken from wearable technologies (smart phones and wrist activity monitors) in the hope of developing systems that streamline referrals to psychiatry.11 As well, the AI-assisted algorithms may provide the basis to develop an early warning system to alert physicians, caregivers and emergency personnel to CLTI patients at risk of suicide / homicide.

In sum, there are multiple new applications for AI, machine learning and deep learning in CLTI. These newer mathematical techniques have the power to surmount hurdles in risk prediction that were previously impossible. Future applications could serve as a decision-making assistant in real-time, as well as improve detection and treatment of depression in CLTI.