Dedicated GenAI Improves Hospital Scheduling

Hospital patient scheduling requires planners to balance staff availability, treatment order, room use and individual requirements while maintaining a workable timetable. A study in Health Informatics Journal evaluates a dedicated generative artificial intelligence system designed to turn natural-language requests into customised scheduling problems. The system imports operational data, interprets planners’ instructions through a neural network and automatically adapts an optimisation algorithm to the stated objectives and constraints. It was tested on a day surgery case from a large public hospital in China, where several types of surgery had to be coordinated across clinical teams and operating rooms. 

Must Read: GenAI System Targets Hospital Patient Scheduling 

Natural-Language Requests Guide Scheduling 

The system brings together a user interface, a neural network parser, an optimisation model, a genetic algorithm and a database. Scheduling data are imported first to create a default problem. Planners then enter requirements in natural language, separating different conditions into short segments. These requests may relate to treatment completion, staff assignments, room availability, deadlines or the order of clinical activities. The parser processes the wording, identifies relevant terms and links them to changes in the default scheduling model.  

A focused neural network is used instead of a general-purpose large language model because the system operates within a defined environment, where users have similar tasks and the range of possible requests is limited. Once the requirements are recognised, predefined code modifications are inserted into the default algorithm. The resulting process checks whether a schedule is feasible and evaluates how well it meets the selected objective. This design allows planners to customise schedules without building mathematical models or writing optimisation code.  

The system can account for restrictions involving specific patients, clinicians, rooms or completion targets, while also changing the performance measure used to judge the timetable. By combining imported operational data with natural-language instructions, the approach aims to make complex scheduling methods more accessible while retaining flexibility for varied clinical requirements. 

Day Surgery Case Tests the Approach 

The evaluation involved 15 patients requiring 3 types of surgery. Each case included anaesthesia, surgery and recovery, with treatment delivered by anaesthesiologists, surgeons and recovery doctors. Operating rooms were allocated according to surgery type, while treatment times varied depending on the clinician assigned. The scheduling task therefore required coordination of staff, rooms and sequential clinical activities under several resource constraints.  

Users provided scheduling requests that were divided into segments to train the parser. The neural network achieved high classification performance on both training and test data, including requests that had not appeared during training. Its average hit rate reached 98% on the test data. After interpreting a user’s instructions, the system created a customised scheduling problem and generated a matching algorithm. In one example, the user requested early completion of all treatments, assigned a patient to a particular anaesthesiologist and limited another clinician’s availability. The system converted these conditions into a tailored problem and produced a corresponding schedule. The algorithm converged after a relatively small number of generations.  

All patient operations were completed within the reported schedule, and the results were displayed as a Gantt chart showing the sequence of activities and their allocation across the operating rooms. The case demonstrated that general scheduling data could be combined with detailed user requirements to produce a complete operational plan. 

Performance Exceeds Current Practices 

The dedicated system was compared with manual scheduling, a rule-based method and branch-and-bound optimisation. It produced a solution more quickly than the comparison methods in the reported example. Its scheduling performance was higher than manual scheduling and the rule-based approach, while remaining close to the branch-and-bound result. Across all user cases, the average difference from branch-and-bound was below 1%, and even the most complex customised problem was modelled and solved within about 20 seconds. The comparison also showed a difference in implementation effort.  

Building and refining the branch-and-bound formulation manually required substantial time, while the dedicated system generated and solved customised problems automatically from planners’ requests. Statistical testing found that its performance was significantly higher than both current practices used for comparison. The findings also indicated that adding more scheduling requirements could narrow the range of feasible solutions and allow the algorithm to converge with fewer iterations. Usability was assessed through feedback from 23 physicians and nurses. The overall score placed the system between good and excellent relative to published scheduling tools. Users responded positively to its reduced complexity, limited cumbersomeness and expected learning time.  

However, staff with little scheduling experience may still find it difficult to express detailed requirements. The system also works as a black box because users do not construct or inspect the underlying optimisation process, which may reduce understanding and trust. Its current evaluation was limited to a single case and 2 constrained resource types. 

A dedicated GenAI system can translate natural-language scheduling requirements into customised hospital optimisation problems and automatically generate algorithms to solve them. In the day surgery case, the parser classified requests accurately, the resulting schedules outperformed current practices and performance remained close to branch-and-bound validation. The system also reduced the time and technical effort needed to formulate complex scheduling problems. Its wider use will depend on whether it can retain this performance in larger hospitals and in settings involving additional resources such as machines or rehabilitation robots. 

Source: Health Informatics Journal 

Image Credit: iStock