Rockwood, Green, and Wilkins' Fractures, 10e Package

127

CHAPTER 5 • Classification of Fracture

classification system is magnified when clinicians use it for dis cussion in real time; this remains a hurdle and raises cause for concern regarding its reproducibility and reliability.

TABLE 5-5. The Orthopaedic Trauma Association Classification of Open Fractures

Skin

1. Can be approximated 2. Cannot be approximated 3. Extensive degloving

ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

Muscle

1. No muscle in area, no appreciable muscle necrosis, some muscle injury with intact muscle function 2. Loss of muscle but the muscle remains functional, some localized necrosis in the zone of injury that requires excision, intact muscle–tendon unit 3. Dead muscle, loss of muscle function, partial or complete compartment excision, complete disruption of a muscle–tendon unit, muscle defect does not approximate

In 1959, Arthur Samuel described AI as “the programming of a digital computer to behave in a way which, if done by human beings, or animals, would be described as involving the process of learning.” 49 As AI algorithms continue to be refined–resulting in a computer’s improved ability to learn without needing to be reprogrammed–its role appears to be increasing throughout nearly every industry and, by extension, in our everyday life. Streaming companies use AI capabilities to make the user’s experiences seamless. By continually feeding data into ever-­ improving algorithms, they gain the ability to parse through an infinite amount of data to create individualized catalogs of music, movies, or TV series. Healthcare is no exception to industry, and the role of AI is increasing in medicine as clinical volumes reach record highs. The use of “Deep Learning” was published in Nature in 2015, and with it, the implications of machine-learning powers skyrocketed. Deep learning allows computational models to learn representations of data consist ing of multiple processing layers with varying levels of abstrac tion and complexity. They describe a method of teaching an AI algorithm classification tasks utilizing deep learning, which turned out to be highly successful at discovering intricacies in complex images. The use of multiple layers allows for detec tion of subtleties in complex imagery by breaking it down into its pixelated values. The key to this system is that it is created from the algorithm’s supervised learning of imagery, rather than a code programmed by human engineers. 34 The use of convolutional neural networks (CNNs) has been touted as the most successful tool for image classification, with significant applications in many domains of science. 34 Their use is now prevalent in biomedical image analysis and has begun to allow for medical image analysis in several fields, including detection of skin cancer, 13 diabetic retinopathy, 20 mammographic lesions, 31 and lung nodules. 25 Unsurprisingly, AI has also been introduced into the practice of orthopaedics. A Korean study in 2018 analyzed 1,891 AP radiographs of the shoulder (1,376 with a proximal humerus fracture and 515 without) with a CNN trained to recognize the presence or absence of a proximal humerus fracture and apply the Neer classification to the fracture pattern fell. The AI model was compared to general physicians, general orthopae dists, and orthopaedists specializing in shoulder surgery. The deep learning CNN model exhibited 99% accuracy in diagnos ing the presence of a proximal humerus fracture and an overall higher accuracy in classification than all groups of physicians, although the difference between the AI group and both groups of orthopaedic surgeons did not reach statistical significance. Specifically, the CNN performed more accurately when com paring three- and four-part fractures; while the human groups

Arterial

1. No injury 2. Artery injury without ischemia 3. Artery injury with distal ischemia

Contamination 1. None or minimal contamination

2. Surface contamination (easily removed not embedded in bone or deep soft tissues) a. Embedded in bone or deep soft tissues b. High-risk environmental conditions (barnyard, fecal, dirty water, etc.) 1. None 2. Bone missing or devascularized but still some contact between proximal and distal fragments 3. Segmental bone loss

Bone loss

Further considerations will be how a summative score may aid in predicting outcome, how the OTA-OFC predicts outcomes in specific anatomic areas that have been understudied (e.g., upper extremity, pelvis), and how the OTA-OFC can reproduce the simplicity of the Gustilo system without losing the increased specificity of the information provided. Goshal et al. investigated the interobserver reliability between the OTA-OFC and the Gustilo classification. 19 They noted sim ilar interobserver agreement between the two classifications, adding that further development was needed to make the OTA OFC a “reliable and robust tool.” 19 Garner et al. used the OTA OFC as an outcome prediction tool in 501 consecutive open tibial shaft fractures and found the muscle component was pre dictive of nonunion, while both the muscle component and the arterial component were predictive of amputation. 17 The reason the Gustilo–Anderson classification remains the standard is perhaps its simplicity. In contrast, the OTA-OFC with its five variables and three levels of severity appears to be too complicated to use in its present form. The challenge of the new OTA-OFC is to demonstrate the additional variable assess ment to be more valuable regarding treatment and prognosis than the size of the laceration and skin envelope alone. The OTA-OFC appears to be a more accurate tool for describing severity of open fractures than the traditional Gustilo–Anderson classification. However, the complexity of the new open fracture