Pharmacovigilance of a high-risk drug: Can AI save us?

Machine learning (ML) is a subdiscipline of AI that uses computerised methods to identify patterns in data to develop a model or algorithm. For drugs like UFH, ML may be used to develop and operate a better method than the nomogram to calculate the best dose for a patient and prevent side effects. Yet, an international review found few ML studies for UFH, none of which had models that were suitable for hospital use.

Thus began a research initiative: to develop an ML model that predicts the blood range from a given dose of UFH. PAH formed a multidisciplinary team of hospital doctors, pharmacists with UFH expertise, and specialists in data management and AI. Over four years, the team gathered nearly 200 patient and drug variables that may influence dosing from approximately 2,500 patients. The resulting model was reasonably accurate, yet, according to local clinicians who tested the model, it was not practical enough to support dosing in hospital patients.

Based on this feedback, our team is undertaking a new project. First, more data has been obtained, now from 4,000 patients across five hospitals, to build a better model. Then the team will reverse the existing model so that it will estimate the optimal UFH dose required to achieve the target blood range, rather than predicting the blood result from a given dose.

The new model will be tested in a silent trial where AI dose estimations will be compared with those by clinicians using the nomogram method. If the ML model method proves to be more effective than the current method, our team plans to put the model in a user-friendly app for all clinicians in the future.

This work is a crucial step towards providing clinicians with real-time recommendations on the best UFH doses in a hospital, and we hope it will ultimately reduce side effects and patient harm. In addition, the AI models have shown us that UFH dosing cannot just be based on patient weight, as many patient and drug factors affect the dose. This is a real advantage of AI.

This research underscores a clear message: more sophisticated dosing strategies are essential to enhance the safety and effectiveness of care. Confronted with a medication whose complexities have challenged clinicians around the world for decades, AI offers a practical pathway to safer, more personalised treatment and a new frontier for pharmacovigilance in everyday clinical practice.