Clinical context

Care pathway

The care pathway for fractures within emergency settings in the NHS is described in the NICE scope1 for this assessment. As part of its assessment, the EAG noted the following additional considerations regarding the care pathway for fractures that was relevant to interpreting the evidence and understanding the potential role of AI as an aid to detecting fractures:

• The scope1 for this assessment was to consider the clinical and cost-effectiveness of AI technologies as a diagnostic aid for detecting fractures in emergency settings, including the emergency department (ED), urgent treatment centre (UTC), and minor injuries units (MIUs). These settings differ in the care pathways available to people with suspected fractures, for example, in the grade of staff who are available to read radiographs and options for further imaging modalities. There may also be differences in the populations who are admitted to each of these settings, for example, MIUs rarely assess people with suspected hip fractures, who in general would be referred to an ED, depending on local policy. In general, MIUs will generally receive a higher case mix of people with suspected fractures of the extremities. Variation in the care pathway and the populations treated within each setting will have implications for the potential value of using AI as a decision aid, as the value of AI is likely to vary according to the staff members using the technology, the fractures assessed, and the downstream impacts of AI on other parts of the care pathway. In order to interpret the evidence for AI as a decision aid presented in this report, it will be important to ensure that the evidence most relevant to the target setting is used and any variation between the evidence base and the target setting is considered carefully.
• Clinical experts noted that the care pathway for the assessment and treatment of fractures varies across the UK as each healthcare service develops and follows its own protocols. While there are established guidelines and standards that guide clinical practice, it is common for services to adapt the care pathway to adjust to the staff and resources they have. This means that there are variations in the staff who provide the initial assessment of X-rays, the use of additional imaging modalities, and the length of time until a definitive report is available. As with the importance of the target setting for the technology, the EAG also noted that the generalisability of the evidence for AI in this report to each healthcare service will vary according to local protocols for assessment and management. In addition to the issues discussed in the NICE scope,1 the EAG also highlight the following considerations:
  • While NICE recommends5 that a radiologist, radiographer or other trained reporter should provide a definitive report of an X-ray prior to discharge, the EAG was advised that this was very rarely possible and the typical time to a definitive report was between 24 hours and 2 weeks across different services. This is much longer than targets set by NHS England who have suggested, in consultation with The Royal College of Radiologists and The Society of Radiographers,6 a 12-hour target to a definitive report of X-rays for outpatients in emergency settings, with the aim that this should be reduced to 4 hours. This will be very difficult to achieve without major changes to the service.
  • The EAG received advice that services also vary in their approach to producing definitive reports, with some centres processing these on a first come, first served basis, while other centres may use alternative strategies, such as only providing a definitive report of X-rays where the initial assessment was negative (i.e. to confirm that a fracture had not been missed).
  • A definitive report from a consultant radiologist or reporting radiographer may not always be necessary depending on the reader of the initial X-ray; for example, a consultant hip surgeon may make the initial diagnosis without the need for a further report.
  • The EAG was aware that centres may take precautions to reduce the risk of missed fractures, particularly for those fractures that are known to be challenging to detect on X-ray. For example, suspected scaphoid fractures and intra-articular fractures may have long-term consequences for a person’s health and mobility if missed, and so the injury may still be treated as a fracture and re-assessed in 2 weeks. A risk-averse approach may also be taken with more vulnerable patients for whom the potential consequences of a missed fracture would be considered to be greatest (e.g. children, people identified as frail). Such precautionary tactics are associated with a reduced risk of missed fractures but an increased risk of overtreatment, including the need for additional assessments after 2 weeks. This means that the potential value of AI as a decision aid will vary according to whether centres use any precautionary tactics.
• There are variations in the staff that would be involved in diagnosing fractures for people with suspected fractures that present to emergency settings out of hours (weekday evenings after 6 p.m., overnight and weekends). During out of hours, some centres may outsource diagnosis of X-rays to centres overseas (i.e. where daytime staff are available due to the time difference7,8). Clinical experts suggested outsourcing diagnosis may not necessarily be as accurate, though the EAG did not have any data to confirm this. There may also be variations in the types of fractures seen in emergency settings during out of hours, due to variation in the cause of injuries; for example, sports injuries are more prevalent at weekends, while alcohol induced injuries (falls and injuries due to violence) may be more prevalent overnight. The EAG therefore considered it plausible that the potential value of AI as a diagnostic aid may vary according to whether people with suspected injuries were presenting to emergency settings during out of hours services or during weekdays.
• The EAG considered it plausible that the introduction of AI as a decision aid within the NHS may lead to broader changes to the care pathway, which may vary across centres and be challenging to predict. For example, centres that use AI may alter the staff that are required to provide the initial assessment, may change their practices for ordering and timing the definitive report, and may alter their use of precautionary tactics. Further consideration of issues related to the introduction of AI within a healthcare setting as discussed in Considerations for implementing AI as a diagnostic aid in clinical practice.

Considerations for implementing artificial intelligence as a diagnostic aid in clinical practice

• As discussed in the review protocol,2 the EAG received advice that AI would likely be valuable for use as a decision aid in a select group of fractures only, such as those that are challenging to diagnose on X-ray during the initial review and where the consequences for a missed fracture are greatest. However, the EAG was uncertain to what extent the technologies could be targeted towards specific fracture sites only or whether, once implemented, the technology would provide an analysis of all X-rays that are entered into the radiology picture archiving and communication systems (PACS). Information provided by the companies involved in this assessment was that services are charged on a fee for scan basis, with some offering volume-based discounts, meaning that the cost of the technology would vary considerably according to whether the technology is used for some or all suspected fractures. If a service chose to target the technology towards specific fracture types and locations, the EAG was uncertain how feasible this would be to implement in practice, for example, whether the technology could be specified to not produce an analysis of certain body locations (as is the case for body locations not covered by the technology) or whether the treating clinician would need to choose to activate the technology for every exam. Expert advice to the EAG was that either approach could be feasible, with the selective use of the technology ensuring that the technology was only used in the minority of circumstances where additional review was beneficial. Cost implications for the technology will vary according to the way in which the technology is implemented.
• The EAG considered that the successful integration of the technology with Radiology PACS, and the perceived ease and acceptability of this, would be important for considering the potential value of the technology. This may include consideration of the effort required to produce a result, the notifications used, and the presentation of the findings. The EAG also considered that, were the technology used, in future it may be required to sit alongside other AI technologies analysing the same images for other anomalies. The EAG was uncertain to what extent the technologies already have this functionality but were aware that other technologies for analysing X-rays for other abnormalities are available.9
• The EAG considered it plausible that the successful integration of the technology and the extent to which staff have confidence in it may influence the way it is used in clinical practice and, therefore, its potential value for the service. For example, where there are discrepancies in the result given by the technology and clinical judgement, staff confidence in the technology may determine which result they choose to prioritise in the final decision. The EAG received advice from clinical experts that confidence in the technology may vary according to staff grade, in that more junior staff may be more likely to place greater weight on the result from the AI. The EAG was also advised that overconfidence in the AI could have negative consequences; for example, staff may rely on the AI diagnosis during busy periods.
• Clinical advice to the EAG was that the optimum order of use would be for the reader to form a judgement about the X-ray without use of the technology, and only then consult the results provided by the technology. This was partly to avoid over-reliance on the technology, with another benefit being to reduce the risk that readers become less skilled in reading X-rays over time, with potential knock-on consequences (e.g. were the technology to not be available in the future). Were the technology to be used in this order, this would increase the reporting time required for X-rays, with potential knock-on consequences for the service.
• The algorithm within AI technologies used for detecting fractures specifies the threshold at which an identified anomaly is defined as a fracture or not, that is the level of confidence required for the technology to return a positive result. Depending on the threshold used, this may favour sensitivity or specificity of the technology, according to whether the threshold is selected to prioritise to avoid missed fractures or FPs. The EAG noted the following considerations on this topic:
  • Generally speaking, the preferred threshold may vary according to the fracture being assessed or the needs of the patient. For example, a lower confidence threshold may be chosen for scaphoid fractures when there are significant consequences to the patient of a missed fracture. The EAG was uncertain whether it was possible for the operator to adjust the threshold used by the technology according to the fracture being assessed or whether this changed manually. Based on the information received, the EAG considered it more plausible that the same threshold was used for all fractures. As a consequence, the EAG considered it important that instructions and training for operators encourages the use of clinical judgement to interpret binary responses from technologies.
  • Some technologies, particularly more recent versions, include a note about the confidence of the result, such as by using a boundary box around the identified anomaly. The EAG considered that confidence metrics may be interpreted differently across users, although this was something that it was not possible to evaluate within this EVA.
  • There is no clear label or metric to identify the threshold used by the technologies during the published evaluations (e.g. there is no apparent scale from which to report the threshold used and compare this across studies and different technologies). This makes it difficult to evaluate the performance of different thresholds, and it was also not clear to the EAG when studies of the same technology were using the same or different thresholds. As the sensitivity and specificity of the technology for detecting fractures will vary according to the threshold used, this creates significant uncertainty in the generalisability of the evidence base to clinical practice.
• AI technologies may be updated over time to refine and improve the technology. These iterative updates may lead to dynamic changes to the cost-effectiveness of the technology and it was not clear to what extent NICE or local services could monitor this. Expert advice to the EAG was that frequent updates to the technology may be challenging for UK bodies to appraise, and it may be that an approved AI technology would not be expected to undergo update except to correct any defects or safety considerations.
• The EAG was unclear about any legal or ethical implications around using the technology within the NHS. For example, how the introduction of the technology would affect liability considerations for missed fractures discussed in the NICE scope.1

Equality issues

Equality considerations for this assessment were noted in the NICE scope.1 Further to these, the EAG identified a paucity of evidence for particularly populations who may be more vulnerable to missed fractures and overtreatment, including people with frailty and those with health conditions that affect their bone health and long-term recovery.