JMIR Publications

ABSTRACT

Background:

Machine-learning-enabled prognostic models are increasingly proposed to support surgical decision-making for degenerative lumbar disorders, yet their clinical adoption remains limited. Understanding how surgeons perceive these tools is critical for effective implementation, particularly given the high-stakes nature of spine surgery where decisions carry long-term functional consequences and medico-legal implications.

Objective:

This study aimed to explore how consultant-level spine surgeons perceive machine-learning-based prognostic tools, including their trustworthiness, clinical utility, usability, workflow integration, and implications for patient counselling and shared decision-making.

Methods:

A qualitative study using semi-structured one-to-one interviews was conducted with 11 consultant-level orthopaedic and neurosurgeons practising in Singapore (response rate: 73% of 15 invited). Participants had a mean of 10.9 years (range: 8-25 years) of spine surgery experience; 82% (n=9) were male. Interviews (range: 15-57 minutes) were transcribed verbatim and analysed using Braun and Clarke's six-step reflexive thematic analysis. Data collection continued until information power was achieved, with three additional interviews completed after thematic sufficiency was reached as participants had already consented. The study was designed and reported in accordance with COREQ criteria.

Results:

Three overarching themes were developed. Trust contingent on data integrity revealed that surgeons' confidence depended fundamentally on data quality, local representativeness, labelling credibility, and rigorous validation, with participants consistently emphasising population representativeness as a trust prerequisite. Surgeons operationalised trust through accessible performance metrics, with most identifying AUROC as their preferred credibility heuristic. Pragmatic orientation as important for implementation demonstrated that usability and seamless electronic health record integration were non-negotiable prerequisites, with surgeons explicitly stating that manual data entry would preclude adoption. Medico-legal concerns were prominent, with participants emphasising that decision responsibility remains with the clinician. Surgical decision-making as a delicate dance between art and science reflected how prognostic outputs were positioned as adjunctive inputs to be reconciled with experiential judgement and patient heterogeneity. Surgeons emphasised that identical prognostic information would be interpreted differently based on practice philosophy, and most highlighted the inherent divergence between technical success and patient-perceived success, underscoring prognostic tools' value for expectation management rather than deterministic prediction.

Conclusions:

This first qualitative study of spine surgeons' perceptions reveals that adoption of machine-learning-based prognostic tools is contingent on data integrity, pragmatic workflow integration, and alignment with professional judgement and not predictive performance alone. Surgeons expressed cautious openness, viewing these tools as valuable in complex cases for clarifying outcome expectations without displacing clinical responsibility. Meaningful implementation requires robust data governance, contextually grounded validation, seamless electronic integration, and explicit positioning of machine learning as supportive rather than substitutive of surgical judgement. These findings provide empirically grounded guidance for developing clinically acceptable and implementation-ready prognostic decision support systems.