Artificial Intelligence for the Orthopaedic Surgeon

An Overview of Potential Benefits, Limitations, and Clinical Applications

Eric C. Makhni, MD, MBA; Sonya Makhni, MD, MBA; Prem N. Ramkumar, MD, MBA


J Am Acad Orthop Surg. 2021;29(6):235-243. 

In This Article


Broadly speaking, the term AI (Table 1) refers to the "mimicking of human cognition by computers."[4] One important subset of AI is that of machine learning (ML), which involves the use of computational algorithms that can analyze large data sets to classify, predict, or gain useful inference.[5,6] In its most rudimentary form, ML models are given inputs and outputs of "training sets" using real-world data to analyze and determine relationships using various methods of pattern recognition. The models are then tasked with creating predictions, given inputs from a "testing set," and these predictions are compared with actual known outcomes to quantify and refine the accuracy of the algorithm with positive or negative reinforcement. These algorithms are comparable to the same experiential "learning" associated with human intelligence, having the capacity to continually assess, and improve the quality of its analyses, given an adequate amount of data inputs, with the potential to continue learning after implementation because new data are available.[7–9] Thus, the predictive power and accuracy of an AI algorithm is only as powerful as its training experience and volume, not unlike the expertise and judgment of a surgeon. Moreover, these algorithms can be seen as doing similar tasks as traditional regression analyses, which determine relationships between disparate variables.

Deep learning can be thought of as an additional subset of ML. Made possible with increasingly powerful computational processing capabilities, deep learning models are more sophisticated algorithms that require less human supervision for development. Also known as deep neural networks, these models mimic the structure and function of the neuron by receiving several inputs (ie, dendrites) than determining which signal meets the internal threshold to pass forward along the axon to the next neuron. Unlike traditional ML algorithms, which generally require human expertise and the predetermined transformation of raw data into a suitable format, deep learning models are a form of representation learning. They function autonomously, allowing the system to discover alternative representations with differing levels of abstractions. The neural network begins with an input tier that receives the raw data. The network then progresses to a number of "hidden tiers" that each respond to different features of the input.[10] Similarly, "backpropagation" of the neural network exists, in which the model continues to learn through the refinement of weighting regarding the known training sets. Through this process of developing multiple hidden layers, the model continues to develop more and more abstract representations of the data. Similar to the way the human brain functions, the machine is able to make "neuronal" connections from "dendrites" on multiple hierarchical data levels.[8] Eventually, the model learns to appreciate a concept on multiple layers and dimensions, building on itself to create a web of interconnected relationships.[7] However, there is room for bias because these models rely on arbitrary weightings that must be manually assigned.