FOUNDERS' AWARD PAPER: Machine Learning Analysis of Multispectral Imaging and Clinical Risk Factors to Predict Amputation Wound Healing
John J. Squiers1, David Bastawros1, Andrew J. Applewhite2, Ronald D. Baxter1, Faliu Yi3, Peiran Quan3, Shuai Yu3, Jeffrey E. Thatcher3, J. Michael DiMaio1, Dennis R. Gable1
1Baylor Scott & White The Heart Hospital, Plano, TX, 2Baylor University Medical Center, Dallas, TX, 3Spectral MD, Dallas, TX
Over 150,000 patients undergo non-traumatic lower extremity amputations in the United States annually.1 Selection of amputation level is a complicated clinical decision. The surgeon must balance the most functional level of amputation to maintain or improve quality of life against the likelihood for failure of healing at any given level of amputation (LOA). No single test is currently accepted as the gold-standard for prediction of amputation wound healing.2 Surgeons must determine optimal LOA using clinical judgement, combining patient clinical risk factors, physical exam, and any available invasive or non-invasive vascular studies. Up to one-third of amputations fail to heal, requiring re-amputation to a more proximal level.3 Failure of primary amputation wound healing decreases patient quality of life and adds substantially to the costs of care for patients with critical limb ischemia.4 Herein, we report the results of a pilot study evaluating the performance of an imaging system that implements a machine-learning algorithm to evaluate multispectral images of superficial tissues at selected LOA combined with patient clinical risk factors to predict amputation wound healing.
This multicenter, IRB-approved, prospective study enrolled subjects with critical limb ischemia planned for lower extremity amputation. Adult patients were eligible for enrollment if they were planned for amputation due to critical limb ischemia with no option for revascularization and had an anticipated life-expectancy exceeding 3 months. Patients scheduled for any procedure other than amputation which may affect wound healing (including invasive angiography) within the 30-days following amputation were excluded.
Prior to imaging, the surgeon declared the intended LOA using clinical judgement based on subject history, physical exam, and any available perfusion studies such as invasive or non-invasive angiography, ankle-brachial indices, or toe pressures. Multispectral images of the superficial tissues at the intended LOA were acquired circumferentially to ensure inclusion of the amputation wound skin flap no more than 14 days prior to amputation (Figure 1). Surgeons were blinded to the multispectral imaging data. Subject past medical history including commonly accepted risk factors affecting wound healing were prospectively recorded.
Post-amputation wound care was performed according to the preferences of the operating surgeon. Subjects underwent a standardized wound healing assessment on approximately postoperative day 30. The surgeon, still blinded to multispectral imaging data, graded the amputation wound as “healing” or “non-healing” according to prespecified criteria. Non-healing status was assigned to any amputation wounds with necrosis, infection, ulceration, dehiscence, or hematoma formation; healing status was assigned to wounds with re-epithelialized tissue at the incision site with total absence of any criteria for non-healing status.
The multispectral imaging device measured reflectance of 8 wavelengths of light in the visible and near-infrared spectrum (400 - 1,000 nanometers) to quantify key superficial tissue properties relevant to the microcirculation.5 A commonly used machine-learning architecture for image segmentation (Very Deep Convolutional Networks for Large-Scale Image Recognition6) was modified with a Feature-wise Linear Modulation technique7 to integrate multispectral imaging data with patient clinical risk factors in order to train a machine-learning algorithm to predict primary amputation wound healing. The included patient clinical risk factors were determined by ranking absolute correlation coefficients between the prespecified clinical risk factors and 30-day wound healing outcome, utilizing clinical risk factors with an absolute correlation greater than 0.25.The algorithm was trained and tested using leave-one-out cross-validation.8
The primary outcome was the performance of the machine learning algorithm using multispectral imaging and patient clinical risk factors as inputs to predict primary amputation wound healing. Secondary outcomes of interest were algorithm performance using only multispectral imaging data or only patient clinical risk factor data as inputs. Algorithm performance was assessed with measurements of accuracy, sensitivity, specificity, and area under the curve (AUC).
Twenty-six patients were enrolled: 22 subjects completed the study and 4 were lost to early mortality. Twenty-five primary amputations were performed (10 toe, 5 transmetatarsal, 8 below-knee, and 2 above-knee). Of these, 11 (44%) were non-healing after standardized assessment (Figure 2). Surgeon judgement was 56% accurate for predicting overall primary amputation wound healing: 40% accuracy for minor amputations (toe/transmetatarsal) and 80% accuracy for major amputations (below/above-knee). Patient demographics and clinical risk factors are listed in Table 1; the 11 bolded entries had absolute correlation with wound healing exceeding 0.25 and were included in the machine-learning algorithm. When both multispectral imaging data and patient clinical risk factors were included in algorithm training, performance was 88% accuracy, 91% sensitivity and 86% specificity (AUC 0.89). The algorithms trained on multispectral imaging data alone (71% accuracy, 66% sensitivity, and 75% specificity [AUC 0.70]) and clinical risk factors alone (70% accuracy, 46% sensitivity, 84% specificity [AUC 0.65]) underperformed as compared to the algorithm combining both data sets (Table 2).
In this pilot study, we have demonstrated the potential utility of an imaging system that combines multispectral imaging, patient clinical risk factors, and machine-learning analysis to predict primary amputation wound healing. The imaging system outperformed surgeon judgment (88% vs 56% accuracy).Amputations due to critical limb ischemia place a significant burden on our healthcare system, and re-amputation/re-operation that occur after failure to heal the primary amputation significantly increase costs and hinder patient quality of life.4 Surgeons must use their best judgement, considering patient clinical risk factors, physical exam, and potentially a variety of invasive and non-invasive vascular studies to select optimal LOA. In order to offer the best functional outcome, surgeons must also balance a desire for maximum limb-salvage against the increased likelihood that primary amputation wounds will fail to heal at more distal LOAs. Unfortunately, there is currently no available technology that is widely accepted and validated to select optimal LOA. This information and technology gap was recently highlighted by the American Heart Association Council on Peripheral Vascular Disease, which released a scientific statement summarizing the limitations of current technology and advocating for the development of improved imaging technologies.2
Historically, ankle-brachial indices (ABIs) have been considered when selecting LOA.9 However, the limitations of ABI for prediction of amputation healing have been repeatedly demonstrated.10,11 Furthermore, nearly 20% of patients have non-compressible vessels, rendering ABIs non-diagnostic.12 Other technologies, including transcutaneous oxygen measurement (TCOM) and laser Doppler imaging (LDI), have been thoroughly studied for predication of amputation wound healing.13,14 Neither have become widely adopted due to important limitations including high intra-operative variability, sensitivity to motion artifact and ambient room temperature, and uncertainty about interpretation of data output thresholds to predict healing.2,15,16
The imaging system investigated in this study may overcome several limitations of TCOM and LDI. The system combines imaging data with patient clinical risk factors that influence wound healing, rather than focusing on a single measurement in isolation from other important conditions that may affect wound healing. The algorithm was trained to provide an easily-interpretable, binary output - healing versus non-healing - rather than complex imaging maps or indices that may be subject to variable interpretation. There are also several limitations. This is a small pilot study. Formal algorithm training must be performed with a larger sample size, and validation of the algorithm should occur in an independent data set before the wound imaging system should be used to aid in clinical decision making. The performance of the wound imaging system may also be subject to other factors not considered in our study including ambient room temperature during image collection as well as patient skin tone.Up to one-third of primary amputations require revision to a more proximal level.3
There is no gold-standard test currently available to predict healing of an amputation wound with high reliability. In this pilot study, we have demonstrated proof-of-concept for a novel imaging system that combines multispectral superficial tissue imaging with patient clinical risk factors via machine-learning analysis to predict primary amputation wound healing. Larger algorithm training and validation studies are planned. If the validation studies confirm our findings, this imaging system may offer significant savings and improved quality of life for patients with critical limb ischemia undergoing lower extremity amputation.
1Geiss LS. Diabetes Care 2019;42:50-4. 2Misra S. Circulation 2019;140:e657-72. 3Kono Y. Ann Vasc Surg 2012;26:1120-6. 4Dillingham T. Arch Phys Med Rehabil 2005;86:480-6. 5Zonios G. J Invest Derm 2001;117:1452-7. 6Simonyan K. arXiv 2014;1409.1556v6: 1-14. 7Perez E. arXiv 2017;1709.07871v2:1-13. 8Hastie T. The elements of statistical learning. 2001. 9Aboyans V. Circulation 2012;126:2890-909. 10Sukul D. JACC Cardiovasc Interv 2017;10:2307-16. 11Bunte MC. Vasc Med 2015;20:23-9. 12Arain FA. J Am Coll Cardiol 2012;59:400-7. 13Keyzer-Dekker C. J Wound Care 2006;15:27-30. 14Mars M. Eur J Vasc Endovasc Surg 1998;16:53-8. 15Goodall R. J Vasc Surg 2019;69:315-7. 16Faglia E. Eur J Vasc Endovasc Surg 2007;33:731-6.
Back to 2021 Abstracts