Rostering AI - time for a rethink?
We wrote a post about the shortcomings of AI for rostering way back in 2018, when the author thought it important to explain that AI meant ‘artificial intelligence’. How things have changed.
What has not changed are the key points in the last article about the magnitude of the challenge: hospitals are complex places, doctors are not interchangeable, hospital resources are limited, and the rewards for effort are sometimes small. And in fact, HosPortal’s overall approach has not really changed that much either: automate where it makes sense, provide powerful tools to the roster administrator, and provide the right balance of guiding rules and flexibility when building rosters.
But the things that are different from 2018 are not what you might think. Generative AI tools that have earned a massive amount of publicity in the last 12 months, like ChatGPT or Midjourney (that made the images for this blogpost), still require far more data and more training than any real hospital rostering solution could realistically get access to. And a human overlay to assess the quality and accuracy of a roster is still required, which is unlikely to save a huge amount of time when real people build real rosters. Even if the AI companies offered a roster-specific version of their AI tools, which they don’t.
What has changed are HosPortal’s abilities and functions, and some clever thinking about what makes a good roster for administrators and users. Over the last 18 months we have made significant improvements in many areas, such as:
the types and variety of rules that an administrator can specify.
the way users can express preferences for their shifts and working days, and balance positive preferences from block-outs that they cannot do.
ability to tag and identify certain types of activities, and include that categorisation in reporting.
ability to specify staffing patterns (such as 3 night shifts followed by 2 evening shifts followed by 4 days off) and term structures.
The bit that really puts the icing on the cake is something we are working on now. Rather than rely just on newfangled AI (although there is a small bit of this, too), we are building on the effort of generations of mathematicians and operations research (OR) academics who have thought long and hard about rostering optimisation, staffing logistics and workforce management. An entry into the academic underpinnings of such thinking, often called the nurse rostering problem or NRP, can be found in this Medium article here.
We hope to have some more news in the next few months about how this will all come together to allow users to combine all the improvements listed above in a seamless way.
Our aspiration is to provide administrators with the tools to automatically collate user preferences and then automatically build a roster using preferred patterns and your local rules in a way that optimises for both the user (‘did I get my preferences?’) and the hospital (‘did I fill my roster within the required working rules?’). We think it will be a game changer and - as far as we can tell - will be better than any other medical rostering tool available.
Stay tuned!
