Artificial intelligence and spine: rise of the machines

Abstract

Machine learning – the data-driven subfield of artificial intelligence – seems like a black box for many of us who are used to interpreting treatment effect sizes and measures of associations using regression models. This issue of The Spine Journal, guest edited by Joseph Schwab, MD, and Aditya Karhade, MD, makes machine learning and other artificial intelligence applications, accessible. In essence, machine learning algorithms are “trained” so they can automatically select and weight factors that maximize classification or predictive power. The prediction models can iteratively be refined to continuously improve with added data and validation. In this issue, we have invited articles from experts in the field to help demystify these approaches as applied to spine research. They provide a taxonomy of machine learning and related fields, [[1]Ghaednia H Fourman MS Lans A Detels KB Lloyd SA Sweeney AA et al.Augmented and virtual reality in spine surgery, current applications and future potentials.SpineJ. 2021; 21: 1617-1625Abstract Full Text Full Text PDF Scopus (6) Google Scholar, [2]Beam A. Sharpening the resolution of data matters: a brief roadmap for understanding deep learning for medical data.SpineJ. 2021; 21: 1606-1609Google Scholar, [3]Joshi RS Lau D Ames CP Artificial intelligence for adult spine deformity: current state and future directions.SpineJ. 2021; 21: 1626-1634Scopus (4) Google Scholar] explain the computer hardware and software requirements[[1]Ghaednia H Fourman MS Lans A Detels KB Lloyd SA Sweeney AA et al.Augmented and virtual reality in spine surgery, current applications and future potentials.SpineJ. 2021; 21: 1617-1625Abstract Full Text Full Text PDF Scopus (6) Google Scholar] and describe their evaluation in terms of calibration, discrimination, validation, reproducibility and generalizability [[4]Azad TD Ehresman J Ahmed AK Staartjes VE Lubelski D Steinen MN et al.Fostering reproducibility and generalizability in machine learning for clinical prediction modeling in spine surgery.SpineJ. 2021; 21: 1610-1616Abstract Full Text Full Text PDF Scopus (6) Google Scholar, [5]Shah AA Karhade AV Park HY Sheppard WL Macyszn LJ Everson RG et al.Updated external validation of the SORG machine learning algorithm for prediction of ninety-day and one-year mortality after surgery for spinal metastasis.SpineJ. 2021; 21: 1679-1686Scopus (3) Google Scholar]. Summarily, Vickers [[6]Vickers A. Decision curve analysis to evaluate the clinical benefit of prediction models.SpineJ. 2021; 21: 1643-1648Abstract Full Text Full Text PDF Scopus (3) Google Scholar] illustrates a simple method to demonstrate when a prediction model can empirically improve the net benefit for a clinical or policy decision. A key distinction is that while traditional regression techniques are well-suited for assessing causation and association, they are poorly optimized for prediction. Machine learning specifically maximizes predictive performance. The approaches are beneficial when confronted with a need to classify high dimensional data where we have more factors than observations (eg, imaging pixels or genomic data,) conceptual uncertainty about what factors are most important for predicting an outcome and a high degree of complex interactions. The use of machine learning is demonstrated in predicting the likelihood of complications and responses following surgery for adult spinal deformity [[3]Joshi RS Lau D Ames CP Artificial intelligence for adult spine deformity: current state and future directions.SpineJ. 2021; 21: 1626-1634Scopus (4) Google Scholar], predicting health-related quality-of-life after surgery for mild degenerative cervical myelopathy[[7]Khan O Badhiwala JH Witiw CD Wilson JR Fehlings MG. Machine learning algorithms for prediction of health-related quality-of-life after surgery for mild degenerative cervical myelopathy.SpineJ. 2021; 21: 1659-1669Abstract Full Text Full Text PDF Scopus (9) Google Scholar] and predicting long-term opioid use in patients undergoing tumor resection [[8]Jin MC Ho AL Feng AY Zhang Y Staartjees VE Steinen MN et al.Predictive modeling of long-term opioid and benzodiazepine use after intradural tumor resection.SpineJ. 2021; 21: 1687-1699Scopus (1) Google Scholar]. Related methods, such as natural language processing of free-text operative reports appear to be better at capturing surgical complications than using administrative claims [[9]Karhade AV Bongers M ER Groot OQ Cha TD Doorly TP Fogel HA et al.Development of machine learning and natural language processing algorithms for preoperative prediction and automated identification of intraoperative vascular injury in anterior lumbar spine surgery.SpineJ. 2021; 21: 1635-1642Abstract Full Text Full Text PDF Scopus (10) Google Scholar]. The Skeletal Oncology Research Group (SORG) ML-algorithm presented is an illustrative application [[10]Karhade AV Thio QCBS Ogink PT Shah AA Bono CM Oh KS et al.Development of machine learning algorithms for prediction of 30-day mortality after surgery for spinal metastasis.Neurosurgery. 2019; 85: E83-E91Crossref PubMed Scopus (41) Google Scholar]. The authors used machine learning to predict mortality in patients undergoing surgery for spinal metastasis disease. A predictive model in this context is important when counseling patients about the decision to undergo surgery. The studies by Shah et al [[5]Shah AA Karhade AV Park HY Sheppard WL Macyszn LJ Everson RG et al.Updated external validation of the SORG machine learning algorithm for prediction of ninety-day and one-year mortality after surgery for spinal metastasis.SpineJ. 2021; 21: 1679-1686Scopus (3) Google Scholar] and Yang et al [[11]Yang J Chen C Fourman MS Bongers MER Karhade AV Groot OQ et al.International external validation of the SORG machine learning algorithms for predicting 90-day and 1-year survival of patients with spine metastases using a Taiwanese cohort.SpineJ. 2021; 21: 1670-1678Scopus (6) Google Scholar] demonstrate external validation of SORG ML-algorithm in distinct study populations, showing excellent discrimination and calibration, and its use improves the net benefit over default strategies applied to all patients or no patients. Machine learning is not without challenges, however. So far, most applications of machine learning in the spine field suffer from a lack of broad external validation and there are only limited applications of their use in predicting patient-reported outcomes, as represented in the study by Khan et al [[7]Khan O Badhiwala JH Witiw CD Wilson JR Fehlings MG. Machine learning algorithms for prediction of health-related quality-of-life after surgery for mild degenerative cervical myelopathy.SpineJ. 2021; 21: 1659-1669Abstract Full Text Full Text PDF Scopus (9) Google Scholar]. Data leakage, overfitting and limited transparency can lead us to errors and mistrust [[4]Azad TD Ehresman J Ahmed AK Staartjes VE Lubelski D Steinen MN et al.Fostering reproducibility and generalizability in machine learning for clinical prediction modeling in spine surgery.SpineJ. 2021; 21: 1610-1616Abstract Full Text Full Text PDF Scopus (6) Google Scholar]. Most spine applications for machine learning only present predictions of an outcome conditioned for a given treatment rather than in the context of a comparative effectiveness treatment decision (eg, the decision between operative versus nonoperative care for disc herniation or between decompression with or without fusion for grade I degenerative spondylolisthesis). Perhaps more troublesome is the concern that training an algorithm can amplify existing social biases, such as systematically underestimating illness severity in Black patients [[12]Obermeyer Z Powers B Vogeli C Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations.Science. 2019; 366: 447-453Crossref PubMed Scopus (623) Google Scholar]. Developers of machine learning algorithms must therefore be mindful of these concerns, be transparent in the analyses, and state interpretations within the limits of the underlying data-generating design. Given an exponential increase in health care data collection and growth of “value-based” reimbursement policies, the interest in prediction models for spine outcomes will likely grow. However, their implementation into routine clinical practice must not be a burden. Karhade et al [[13]Karhade AV Schwab JH Fiol GD Kawamoto K. SMART on FHIR in spine: integrating clinical prediction models into electronic health records for precision medicine at the point of care.SpineJ. 2021; 21: 1649-1651Abstract Full Text Full Text PDF Scopus (1) Google Scholar] suggest using the digital SMALT on FHIR standard as a forward direction to integrate machine learning algorithms into an efficient clinical pathway, potentially fostering their dissemination and adoption. The exciting methodological innovations of machine learning in back pain hold great potential for improved variable selection, outcome identification, prediction and decision-making. We hope these approaches will eventually lead to improved medical decision-making and better outcomes. The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.