The challenge

The identified challenge was the lack of a structured, reusable approach to preparing patient safety discussions for M&M conferences. The aim was not to automate clinical judgement, but to test whether a model-based risk analysis derived from team knowledge could serve as a structured input for drafting an M&M decision template.

M&M preparation often relies on fragmented information and informal interpretation. In complex clinical environments, such as coronary angiography, risks do not arise from a single isolated factor. They emerge from the interaction between tasks, people, technology, information flow and organisational conditions.

In this specific pilot example, the safety concern was a risk scenario in coronary angiography in which cognitive overload during real-time decision-making and escalation could contribute to complications not being detected in time. This formed the basis for testing whether a structured model could provide a clearer and more traceable starting point for discussion.

Method and measures

To explore this, a systems model based on Systems Engineering Initiative for Patient Safety (SEIPS) 2.0 was created in Systems Modeling Language (SysML) using SPARX Enterprise Architect. The objective was to represent the work system, the contributory task factor, the resulting risk and the proposed measures in a traceable form.

The model focused on one coronary angiography scenario. The critical task factor was described as cognitive density in real-time decision-making and potential escalation. In the model, this contributed to the risk that complications would not be detected in time. The text states an impact on quality of care, an occurrence rating described as relevant and an overall risk class of moderate.

The proposed measures were:

• pre-procedure briefing
• risk-adapted staffing
• standardised laboratory layout
• regular simulation drills.

The intended achievement was a more structured, transparent and reusable basis for M&M preparation and discussion.

Outcomes and lessons learned

The pilot showed that a structured model can be a useful way to organise safety-relevant knowledge. Because the model linked work system elements, risks and measures in a traceable way, it provided a clearer starting point for discussion than unstructured text alone.

The practical process tested in this pilot was:

• defining a relevant patient safety scenario in coronary angiography
• modelling the work system and the contributory task factor
• linking this to a patient safety risk
• documenting possible mitigating measures
• using the model as the basis for an AI-assisted one-page decision template.

One important observation was that the AI-generated output reflected the underlying model's content. This suggests that a structured model can support more consistent synthesis than relying only on memory or informal interpretation.

The text does not describe multiple alternative technical approaches in detail, so it cannot be stated from the source whether other options were formally compared or ruled out. It also does not state direct patient involvement. Staff involvement is referenced indirectly by using team knowledge as an input to the model.

The text does not report formal measurement tools, outcome metrics, time savings, patient safety indicators or model costs. Therefore, no validated impact measurement can be claimed from the source.

A key lesson learnt was that AI can assist with drafting and synthesis, but cannot replace clinical judgement, governance or safety review. Any output generated from the model still needs to be checked against the source material and reviewed by responsible clinical and patient safety leads.

Impact

This work is only a prototype, not as a formal effectiveness study. As a result, the impact that can be claimed is limited.

The main result was that the structured model appeared to support:

• clearer organisation of safety-relevant knowledge
• better traceability between work system factors, risks and proposed measures
• a more consistent starting point for multidisciplinary discussion
• reuse of modelled information for drafting a one-page M&M decision template.

At the same time, the the study is explicit about what was not demonstrated. The pilot did not test whether the approach:

• improved patient outcomes
• reduced harm
• shortened preparation time in routine practice
• improved care delivery in a measurable way.

A further limitation was that only a single, limited example was used, and some information was withheld for data protection reasons. This means the results were narrower than would be needed for broader implementation decisions.

What worked was the structured linkage between the work system, contributory factors, risks and measures. What remains uncertain is whether this translates into measurable operational benefit in routine clinical governance.

A likely barrier to improvement is the need for continued expert review, because AI-generated output cannot be used without clinical validation and governance oversight.

If repeated, the next stage would need a clearer evaluation design, including defined measures of clarity, consistency, usability and possibly preparation time.

Next steps

The next step is a practical pilot in real clinical governance settings. A suitable next-stage comparison would be conventional M&M preparation versus model-supported preparation in a small, clearly defined pilot.

The proposed questions for the next phase are:

• Does the approach improve clarity and shared understanding?
• Does it help teams identify contributory factors more systematically?
• Does it support consistency and traceability of measures related to patient safety?

The study does not provide evidence of long-term organisational change, staff reaction, patient impact statistics or system-wide implementation results. Therefore, those elements cannot yet be stated as outcomes.

However, based on insights from the pilot study, the anticipated longer-term value would be to make patient safety knowledge:

• more structured
• more reusable
• easier to discuss across professional groups
• more clearly linked to the wider work system rather than to isolated errors.

A sensible next step would, therefore, be a controlled local test with defined governance, clinical review and evaluation criteria before any broader adoption.

While the regulatory environment for each country is different, some common ingredients for success are emerging. Across these examples we tended to see: 

1. Significant investment made over the years in their data infrastructure.
2. Some kind of innovation centre or hub allowing access by in-house clinicians and scientists, and by vendor partners, to test ideas using patient-level data.
3. A balanced approach to AI development – part in-house, often led by clinicians, and part procured from an AI vendor.
4. A centre or unit focused on AI governance, including standard agreed rules for testing AI in real-world contexts. These focused on going beyond the early-stage development of AI models to investigate how things panned out ‘on the ground’ when implemented.
5. Partnerships with large technology companies, such as Amazon Web Services, Microsoft and Google, stretching over years.
6. Built-in training for staff on how to develop, test and use AI effectively.

Many of these health facilities focused on AI to tackle challenges common to many settings, such as: 

• Improving productivity and releasing clinical capacity, particularly by reducing administrative burden and improving operational efficiencies.
• Reducing waiting times by enabling earlier clinical interventions, streamlining processes and pathways, and speeding up discharge.
• Improving safety and clinical outcomes via predictive analytics to identify high-risk patients or post-operative complications.
• Promoting personalised medicine, combining genomics, imaging and electronic health record data to advance research and provide tailored treatments.