cover story two

Unveiling the Future

Artificial Intelligence in Spine Surgery

By Omar Ramos, MD

Basic Concepts

Proper knowledge of the terminology associated with AI is critical to establishing a good foundation for its understanding. The main goal of AI is to allow computer systems to perform tasks that would normally require human intelligence. These tasks may encompass a broad spectrum of functionalities, including understanding spoken language, recognizing patterns, solving problems, making decisions and executing physical actions. Some of the pillars of AI include Machine Learning (ML), Deep Learning (DL), and Natural Language Processing (NLP).

AI algorithms can also incorporate a rich array of patient-specific variables..

ML is a subset of AI that enables computers to learn from data and improve their performance without being explicitly programmed to do so. It involves developing algorithms and models that can make predictions or take actions based on patterns and examples in the data. In ML models created with supervised learning, the model is trained on a labeled dataset with both input data and the corresponding correct output. The system then learns the underlying pattern and the structure from the data and can analyze new datasets that are not labeled and provide the correct output. ML models can also be created with unsupervised learning (trained only on input data), where the model identifies clusters and patterns within the dataset and can subsequently identify the presence or absence of similar clusters in other datasets. The last way to train an ML model is by reinforcement learning, where the model makes decisions that maximize the “reward” determined by the programmer. Instead of if/then parameters used in supervised and unsupervised learning, reinforcement learning models consider the environment and make decisions that maximize the reward; therefore: different decisions could be made by the system when presented with similar data if the environment changes. In spine surgery, ML models have focused mostly on predictive modeling using supervised learning. Several authors have reported on models that have accurately predicted the need for intraoperative and postoperative blood transfusions in patients undergoing adult spinal deformity surgery, major complications within the early postoperative period, as well as extended length of stay after spine surgery.

DL is a subfield of ML that focuses on the development of artificial neural networks capable of learning and making decisions in a hierarchical manner. Neural networks are algorithms inspired by the structure and function of the human brain’s interconnected network of neurons. Neural networks are designed to recognize complex patterns and relationships in the data. Deep learning has achieved significant advancements in areas such as image and speech recognition. A recent study presented a deep learning model that was able to detect and classify central and lateral recess lumbar stenosis with comparable agreement to radiologists with subspecialty training.

NLP is a branch of AI that focuses on facilitating the interaction between humans and computers using natural human language. The ultimate objective of NLP is to allow computers to read, decipher, understand and make sense of human language in a manner that is valuable for the computer system. In health care, one of the first uses of NLP was in the creation of digital scribes. Initial systems acted as simple recorders of an interaction, transcribing it into written language. Currently, some systems that use artificial intelligence integrated with NLP can not only transcribe the interaction into written language but also organize it into a uniform, well-structured document that the provider can then review and sign.

AI in Spine Surgery

The goal of incorporating AI into spine surgery is to allow surgeons to become more precise and achieve better outcomes. It implies employing very sophisticated systems that can not only help analyze vast amounts of information but also potentially assist in decision-making processes, therefore minimizing human error.

AI algorithms excel in executing automated tasks in image recognition, segmentation and diagnostics. These AI-guided image analyses are broadening their reach in both surgical and non-surgical environments. Echoing the capabilities of technologies like Facebook’s Segment Anything, neural networks like U-Net, and deep learning algorithms like Mask R-CNN (Region-based Convolutional Neural Network), AI-guided image analyses have mastered the fine mapping of spinal elements, including nerves and paraspinal muscles. Mask R-CNN has been used for segmentation and live tracking of instruments within the surgical field. Utilizing such advanced technology enables surgeons to pinpoint the exact locations of spinal and paraspinal structures, bolstering the precision of tool and implant positioning, minimizing radiation exposure, and fine-tuning the planning of surgical trajectories. Better planning and implant positioning can result in improved outcomes, such as prevention of adjacent segment degeneration due to facet-joint violation and decreased risk of implant malposition requiring reoperation.

AI algorithms can also incorporate a rich array of patient-specific variables—including demographics, anthropometrics and medical and surgical history—to adeptly predict the likelihood of a given outcome based on operative or nonoperative interventions. In contrast to the time-intensive and often complex processes involved with conventional predictive analyses, which heavily rely on manually crafted regression models, AI presents a groundbreaking shift in this domain. It stands unparalleled in identifying intricate patterns and trends, both linear and nonlinear, often outpacing the capabilities of traditional regression analyses in terms of accuracy and efficiency. After an AI model is created, trained and tested, a new set of data can be interpreted in seconds. The creation of these predictive models serves a pivotal role in improving decision-making processes, refining patient-selection criteria and facilating personalized counseling for a spectrum of medical and surgical interventions. A growing number of AI algorithms have already demonstrated proficiency in predicting various outcomes in the realm of spine surgery—ranging from projecting the discharge destinations post elective lumbar fusion to estimating risk of re-herniation after lumbar discectomy.

Most of my experience with artificial intelligence has been on spine surgery research. Earlier this year, I coauthored two publications where we used machine learning algorithms to predict short-term outcomes after posterior cervical decompression and instrumented fusion and the prediction of discharge destination following elective lumbar fusion. Our goal is to incorporate image recognition as well as predictive models in a user-friendly interface where providers can input patient-specific variables and obtain patient-specific recommendations regarding potential treatment options, alignment goals, risks and likely outcomes associated with interventions.

Complex challenges, ethical dilemmas and unresolved queries stand before us.

Data collection

How data is collected for AI purposes depends on the type of model and application it is being used for. Systems used intraoperatively for spinal instrumentation use a combination of imaging data obtained before or during surgery as well as “live” data that can be obtained by different means (stereotactic, magnetic, surface anatomy camera and so forth). AI models used in predictive modeling generally utilize data existing in public and commercial health insurance databases, electronic medical records and picture archiving and communication systems (PACS). As a general concept, the larger the dataset that a model trains with, the better predictions it will be able to make. The collection process is dynamic and ever evolving, aiming to include a broader spectrum of data types to develop a more holistic understanding of patient health and surgical outcomes. The potential to integrate genomic data, for instance, might further personalize treatment strategies. 

Though promising, it’s essential to acknowledge that the realm of AI predictive modeling is still in its nascent stages, largely constrained by the current limitations inherent to the datasets employed for model training. The existing lack of depth and granularity within the larger medical databases poses a considerable barrier, restricting the scope and precision of hypothesis testing. Moreover, persistent issues related to data coding discrepancies and standardization gaps have temporarily hindered the pace of algorithmic evolution.

The integration of AI in spine surgery undeniably opens a plethora of avenues that are patient centric. Although medical providers make medical and surgical treatment recommendations based on available literature evidence as well as individual patient factors, AI allows the incorporation of a much larger amount of population data as well as individual patient factors when making these recommendations. The goal is for AI to enhance the ability of the medical provider to make evidence-based decisions that are personalized for each patient with ease. This shift towards a more patient-centric approach also fosters enhanced patient satisfaction, as outcomes are aligned more closely with individual expectations and needs. 

Challenges and the future of AI

Looking ahead, the future of AI in spine surgery seems to be brimming with potential. AI systems are anticipated to become even more intelligent and nuanced, possibly taking on roles such as more extensive robotic assistance during surgeries, virtual postoperative monitoring and utilizing big-data analytics for continuous improvement in patient care.

There are also, however, significant barriers to adoption. These include the high costs associated with implementing AI systems, potential cybersecurity threats and data privacy concerns. Moreover, there is also a need for establishing standardized protocols and guidelines to govern the integration of AI into clinical practice. Now more than ever, a multidisciplinary approach is crucial to ensure that AI technologies in spine surgery are developed, validated and deployed responsibly. The field of Explainable Artificial Intelligence (XAI) seeks to develop models that provide clear explanations of its recommendations, allowing clinicians to understand and validate the rationale behind the AI’s decisions.

Another concern with AI algorithms is the presence of biases, such as racial, cultural and socioeconomical biases. If training datasets contain these biases, they can perpetuate and amplify partiality in clinical decision-making. For example, a recent study published in the World Neurosurgery journal found that minority patients with adult spine diseases were less likely to undergo surgery but more likely to receive surgery from a low-volume provider and experience a higher rate of postoperative complications. If an AI model is trained with a database that includes this data, it may predict that minorities have worse outcomes after surgery, which could lead surgeons to perform even less surgery in minority patients. It is crucial to ensure that the datasets used for training AI models are free from inherent biases. Continuous monitoring and auditing of AI algorithms can help identify and address any biases that may arise during deployment. Developers must also be wary of the ecological fallacy—defined as the act of assuming relationships observed at a group level hold true at an individual level—when deploying and interpreting AI decisions. To foster an equitable and just health care landscape, it is imperative to rigorously scrutinize and curtail biases within the training datasets of AI models. These biases, if left unchecked, can not only perpetuate but also exacerbate discriminatory tendencies in clinical decision-making, consequently undermining the principles of inclusivity and fairness in patient care. At the cornerstone of mitigating this issue is the steadfast commitment to assembling training datasets that are meticulously curated to be devoid of any intrinsic biases. This commitment should be coupled with an ongoing strategy of monitoring and auditing AI algorithms to promptly pinpoint and rectify any biases that might inadvertently surface during operational phases. Furthermore, developers bear the significant responsibility of safeguarding against the ecological fallacy—a logical error where relationships discerned at a collective group level are presumed to be applicable at an individual level. This necessitates a vigilant approach when deploying and interpreting AI decisions, ensuring that the nuanced and unique attributes of individuals are duly considered, fostering a more personalized accurate, and ethical health care decision-making process.

Conclusion

The rapid rise of AI has sparked a whirlwind of excitement and unease in equal measure. While skeptics harbor concerns that AI might overshadow the indispensable role of health care practitioners, it is essential that we, as seasoned professionals in this field, embrace a perspective rooted in enlightenment and collaboration with this burgeoning technology, rather than viewing it as a potential usurper of our roles. A plethora of complex challenges, ethical dilemmas and unresolved queries stand before us. The labyrinthine nature of spine diseases beckons; can the sophisticated algorithms of AI decipher its intricacies step by step, potentially revolutionizing our understanding and management of these conditions? On the flip side, could it eclipse the irreplaceable nuance and empathy that health care practitioners bring to the table, positioning AI as a competitor rather than a collaborative ally? These considerations are not merely speculative but bear a weight of urgency and gravity as we navigate the evolving landscape of health care in the 21st century. It is incumbent upon us to foster a dynamic ongoing dialogue and vigilance, ensuring a thoughtful, balanced integration of AI into our practices, ever mindful of preserving the human touch that lies at the heart of patient-centered care.

Omar Ramos, MD, is a surgeon at Twin Cities Spine Center.