ADOPTing AI to facilitate timely bone care

A practical and clinically-grounded overview of the UK ADOPT (AI-enabled Detection of OsteoPorosis for Treatment) study was presented by lead author Prof Kassim Javaid, Professor of Osteoporosis and Adult Rare Bone Diseases, University of Oxford, UK.

ADOPT uses artificial intelligence (AI) to review hospital computed tomography (CT) scans and identify appropriate patients, who are not critically unwell or in the final stages of life, that will benefit from a prompt bone health assessment. In his talk, Prof Javaid provided insight into AI algorithms in clinical practice, how healthcare systems and clinicians “can use them intelligently”, and why he and his team now include the technology in their osteoporosis pathway.

Prof Javaid began by presenting the empirical data and a patient case study to show that while patients undergo CT scans for many reasons, up to 50 per cent of moderate-to-severe vertebral fractures on images that include the spine are not detected. “This has led to a tsunami of AI models that look at different [radiological] modalities to automate the detection of bone fractures,” he said, adding that healthcare systems “don’t have the manpower” required to do so in the same manner.

Discussing the available technologies, Prof Javaid detailed “ensemble AI” – which involves training several AI models to achieve optimal fracture prediction and detection. “Why use one model when you can use four?” he said. Prof Javaid explained that agreement between the separate models provides “a much higher performance rate” and diminishes dependence on just one system.

“CT has a lot more information than just the shape of the bone,” Prof Javaid continued. He described the “valuable” data AI can produce and analyse regarding muscle, age, and other demographics. Attendees heard that the technology can improve fracture prediction even without analysing radiological images or bone parameters, but instead using data-sets, such as ICD (International Classification of Diseases) 10 codes. He presented evidence to show that AI is “quite impressive” when compared to the traditional mode of osteoporosis detection, the DXA (dual-energy x-ray absorptiometry) scan. Prof Javaid also noted “a massive amount of [AI] models in the literature”. He advised colleagues that they should wait for those systems that are granted regulatory approval, as most will “drop off” during the process.

Practical insight from ADOPT

Prof Javaid proceeded to share some of the practical knowledge he gained while successfully implementing a “very simple AI pathway using CT data” in five hospitals during ADOPT. “We did this in four work packages,” he explained. “Does AI actually work in the hospital setting? What are the regulatory pathways for deployment? Do we actually improve the number of patients we manage? And finally, do we prevent fractures?”

Prof Javaid gave a step-by-step description of the methodology, approval, and implementation phases of the study, as well as an overview of practical issues, tips for adoption, litigation considerations, and “lessons learned”. He emphasised the importance of identifying “humans in the loop”, as well as the “massive IT requirements”. He described carrying out a shadow test before going live “to see what happens” within a clinical context and identify areas to be fixed.

“AI is going to revolutionise healthcare in the coming years,” Prof Javaid told the meeting. “The good news is that after 2,000 patients, we’ve only had positive feedback. Most patients are delighted that AI is now helping them achieve bone health.” Concluding, he emphasised the necessity of “human review”, multidisciplinary collaboration, particularly with general practice, and patient follow-up after the AI has identified those at risk.

“I would suggest you spend 90 per cent of your time thinking about the patient and 10 per cent thinking about AI, because the AI will work. You’ll get lots of patients and your AI implementation group has to focus on what happens once they get confirmed.”