Hierarchical clustering of patient data using artificial intelligence (AI) can identify patterns that may guide preoperative decision-making in adult spinal deformity (ASD) surgery by predicting outcomes and major complications, a collaborative study by the International Spine Study Group and the European Spine Study Group has indicated.
The paper was presented at the Scoliosis Research Society annual meeting (18–21 September, Montreal, Canada) by lead author Christopher Ames (UCSF, San Francisco, USA). Ferran Pellisé (Vall d’Hebron Institute of Research, Barcelona, Spain) presented the findings most recently at the EUROSPINE annual meeting (16–18 October, Helsinki, Finland), where the paper featured in the Best-of-Show session. Miquel Serra-Burriel (UPF, Barcelona, Spain) was the lead AI data scientist for this work.
In the study, Ames and Pellisé note that AI-based, unsupervised clustering provides a wide view of outcomes and complications, and can be used when risk outcome calculators are not feasible or available. Cluster-based risk/outcome classification will likely have a significant impact as a decision support tool, Pellisé said at EUROSPINE. Ames noted at SRS 2019 that the surgeon of the future will be a ‘chimera’ combining human and artificial intelligence to improve patient care.
Through the collaborative analytics study, the US International Spine Study Group (ISSG) and the European Spine Study Group (ESSG) aimed to create a risk-benefit adult spinal deformity (ASD) classification or efficiency grid using AI unsupervised cluster analysis which is intended to assist surgeons and patients to identify treatment options yielding the highest healthrelated quality of life improvement, combined with the lowest risk.
In his presentation, Pellisé noted that there are existing, reliable individual models to predict outcomes, complications, and reintervention following spinal deformity surgery, but that there is a need for computer access to be able to get a prediction. The goal was to have this information at the point of care without the need for computer access.
To build their classification model, the authors analysed data from a cohort of 1,245 ASD patients taken from the ISSG and ESSG databases with baseline, one-year and two-year health-related quality of life and surgical data queried. For unsupervised cluster analysis, the authors fitted dendograms, one with surgical features, and another with patient characteristics. Both were built using Ward distances and optimised with the gap method. Patient characteristics observed included age, sex, height and weight, and number of previous surgeries.
Based on the surgical characteristics, the cluster analysis defined five surgical clusters: short fusion with single posterior column osteotomy; long posterior spinal fusions; multiple posterior column osteotomies; interbody fusion without decompression; interbody fusion with decompression, and patients with three column osteotomies. The analysis identified three clusters based on the patient characteristics: young coronal patients, with a mean age of 36 years and an average mean coronal deformity of 52 degrees; old primary patients, with a 58-year mean age, a mean average of 39 degrees coronal deformity and with no revisions; and old revision patients, with a mean age of 62 years, with larger sagittal deformity and less coronal deformity. Around 80% of patients in the old revision cluster had previously undergone surgery.
These clusters and surgery types were filtered and back-walked to provide decision trees intended to help the user identify which cluster that patient and the surgery belonged to. Normalised two-year improvement and major complications were computed both for the patient and surgery clusters and were used to create an efficiency grid. Pellisé noted that this helps to identify the treatment patterns yielding the optimal calculated quality of life improvement, with the lower risk major complications.
At SRS Ames commented: “This same analysis shows us that contrary to what most people might believe, patients with greater complication risk at the same time have greater benefits, so this is something that we have to take into account. In counterpart, young coronal patients with less risk, they also have less functional gain.” This may have significant implications for how government-based healthcare systems with limited resources spend their healthcare budgets in the future, according to Ames.
As well as helping to aid decision-making preoperatively through predicting outcomes and major complications, the joint ISSG and ESSG study concluded that ASD classification pattern identification could help to facilitate treatment optimisation by educating surgeons on which treatment patterns yield optimal improvement with the lowest risk.