Page 109 - Read Online

Ambati et al. Art Int Surg. 2025;5:53-64 https://dx.doi.org/10.20517/ais.2024.45 Page 55

many recent groundbreaking applications of AI. These models leverage artificial neural networks, which can
learn complex relationships between their complex input data and generate a wide variety of potential
[16]
outputs, including text, images, and audio . Compared to traditional ML, deep learning is more powerful
but requires massively more data and is specialized for data-intensive tasks such as vision and language
processing. On the other hand, traditional ML has straightforward ways for users to understand what
variables are most important in making predictions, which is particularly true for the more simple
methods . Efforts to understand how deep learning makes predictions and how it weighs input data
[15]
[16]
present an evolving challenge .
Another key division among AI/ML algorithms is between supervised and unsupervised approaches.
Supervised models require training between inputs and desired outputs using labeled data. The process of
annotating and curating such datasets can be cumbersome and is particularly challenging in surgical
specialties where patient numbers are low and factors such as patient privacy are essential. Unsupervised
models such as k-means clustering instead can find patterns inherent in unstructured data; however, they
cannot directly make predictions in the same way that supervised algorithms can . While AI algorithms
[15]
are advancing at a staggering pace, developing a general framework that outlines both the capabilities and
limitations of these models will be critical for spine surgeons in the coming decades.

RECENT AI/ML APPLICATIONS IN SPINE SURGERY
Preoperative planning
Spine surgeons face various clinical and radiographic factors in preoperative planning, which they must
parse to make often difficult and subjective decisions about patient selection and surgical approach. AI
could augment the ability of clinicians to understand patient disease states, weigh factors such as symptoms
and disability, and assess anatomic and pathologic parameters from multimodal imaging.

A number of recent studies have used ML to assist in understanding clinical variability and phenotype in
patients undergoing spine surgery. Unsupervised clustering of the clinical metadata of patients with
degenerative spondylolisthesis revealed distinct phenotypes of disease severity, which had different levels of
[17]
postoperative improvement, pain, and satisfaction despite sharing the preoperative severity on imaging .
Clustering has also been shown to disentangle the interactions between patient characteristics and surgical
procedures in adult spinal deformity. By inputting a variety of variables including clinical, disability, and
spinopelvic parameters, the initial clustering algorithm grouped patients based on age and prior surgery,
followed by a second clustering step based on the type of surgery performed, yielding distinct groups that
vary in terms of the risk/benefit of surgery . In spinal deformity, preoperatively oriented algorithms can
[18]
also forecast fine-grained aspects of postoperative responses to a standardized scoliosis questionnaire .
[19]
Together, these studies reveal how AI can uncover patterns in clinical data that could guide preoperative
patient counseling or patient selection and maximize quality and value.

Patient imaging is central to preoperative planning in spine surgery and its analysis is one of the most
promising applications of AI [Figure 1]. Radiographic analysis includes image segmentation, which refers to
the accurate identification and delineation of anatomical structures. Previous work has leveraged deep
learning to segment spinal cord structures in a fully automated manner [20,21] , performing better than previous
state-of-the-art techniques that did not leverage ML . Further studies incorporating data from patients
[22]
with SCI improved on these initial advances. They can capture and identify lesions that correlate with
motor scores at admission and predict thoracolumbar injury classification scores from CT alone, which
[23]
typically require MR imaging to assess ligamentous integrity . Other algorithms can accurately segment
[24]
other relevant anatomic structures, including vertebral bodies and discs [25,26] , as well as paraspinal