Artificial intelligence in orthopaedic oncology Towards algorithm assisted healthcare

Artificial intelligence (AI) encompasses the wide-ranging branch of computer science in which the goal is to replicate or simulate natural intelligence (the intelligent behavior demonstrated by humans) in computers. Machine learning (ML) arose as a subset of AI, where computer systems (or “machines”) automatically improve in identifying patterns from data though experience, rather than being explicitly programmed by humans. A more accurate understanding of the approach, methodology, and pitfalls of developing and using such algorithms is essential in the pursuit to implement algorithms as an aid in daily practice. Therefore, this thesis aims to give a comprehensive overview of the areas where AI is currently being developed and implemented in orthopaedic oncology: ML prediction algorithms, natural language processing (NLP) for automated data collection, and computer vision using Deep Learning (DL) image analysis.
These algorithms were developed, internally validated, or externally validated with novel methods and up-to-date performance metrics. Results of these studies show that ML predication algorithms have superior performance compared to their non-ML counterparts. Repeated external validation of algorithms is different populations is always required. Furthermore, NLP is a viable technique for the automatic identification of adverse events in free text reports, often superior to methods based on diagnosis and procedural codes. Lastly, DL models have comparable performance to clinicians in assessing abnormalities on musculoskeletal images, while diagnostic accuracy considerably improves when clinicians are aided by these models. AI should therefore always be used as a technical supplement in clinical care, not as a replacement for human intelligence.