GenAI System Targets Hospital Patient Scheduling

Patient scheduling in hospitals and clinics remains a complex operational task that affects patient satisfaction and healthcare efficiency. A Health Informatics Journal research article describes a dedicated generative artificial intelligence system designed to handle customised patient scheduling problems through natural-language input. The system imports scheduling data, parses user requirements with a deep neural network and automatically generates a customised genetic algorithm to solve the resulting optimisation problem. It was tested in a real-world day surgery case involving biliary surgery at a large public hospital in China. The results show strong scheduling performance, rapid modelling and higher performance than two current practices.

System Combines Natural-Language Input and Optimisation

The dedicated generative artificial intelligence system consists of a system interface, deep neural network, optimisation model formulator, genetic algorithm and system database. Scheduling data are imported into the system rather than entered through prompts. Planners then enter scheduling requirements in natural language through the interface, allowing them to specify objectives and constraints in detail.

The deep neural network extracts information from each query, including keywords and their relative positions. Users are instructed to divide queries into segments using semicolons. Each segment is processed through lowercase conversion, stemming, lemmatisation and entity linking. The resulting information is used to establish an extended three-field notation for the customised scheduling problem.

The system then translates the identified requirements into modifications of the default patient scheduling model. These modifications can cover objectives, constraints or both. Examples include assigning a patient to a specific doctor, restricting the use of a resource, setting completion targets, minimising the number of patients whose surgery completion times exceed due dates or using one resource more than other resources.

A customised genetic algorithm is automatically generated by replacing or inserting pre-written code into the default scheduling programme. This allows the system to adjust feasibility checks and fitness assessments according to the scheduling requirements entered by the user.

Case Study Uses Day Surgery Scheduling Data

The dedicated system was applied to a real-world day surgery case from the literature. The case involved patients due to undergo biliary surgery at a large public hospital in China. Three surgery types were included: laparoscopic cholecystectomy, cholecystectomy and common bile duct exploration T-tube drainage. Each surgery involved preoperative anaesthesia, surgery and anaesthesia resuscitation.

Must Read: Radiology Scheduling Pressures in German Hospitals

The available clinical staff comprised anaesthesiologists, surgeons and anaesthesia resuscitation doctors. Operating rooms were constrained by surgery type, with four rooms available for each type of surgery. The scheduling data included operation times, operating room constraints and patient data. No missing data were present.

The system identified the default scheduling problem automatically after importing the data. One hundred and one users entered scheduling requirements, which were decomposed into five hundred segments to train the deep neural network parser. Four-fold cross-validation was used. Training was completed in under five seconds on the specified personal computer, and the average hit rates were high for both training and test data.

One example required all treatments to be completed as soon as possible, patient number three to be treated by a specified anaesthesiologist and another anaesthesiologist to be available only after a defined time. The system converted these requirements into a customised scheduling problem and generated code for a genetic algorithm. In this example, the maximum fitness converged after around sixty populations.

Performance Exceeds Current Scheduling Practices

The experimental results showed that all patient operations were completed within the reported scheduling window, with the results presented to the user as a Gantt chart. The system also generated assignments of operations to the operating rooms. The customised genetic algorithm used patient scheduling performance against user expectations as the primary endpoint for the convergence target.

A branch-and-bound algorithm was used to validate the optimal solution for the customised scheduling problem on the same platform. Manual creation and refinement of the customised patient scheduling problem required approximately three hours. The two methods found the same optimal solution in the described comparison, supporting the effectiveness of the dedicated system.

Scheduling results were also compared with manual scheduling and the most-work-remaining rule, with resource constraints such as doctor and operating room availability considered. The dedicated system achieved higher scheduling performance than both current practices and performance close to branch-and-bound. A paired t-test found that its scheduling performance was significantly higher than manual scheduling and the most-work-remaining rule.

Usability testing was conducted with feedback from physicians and nurses using the system usability scale. The results placed usability at an average-to-good level, with particular strengths in reducing perceived complexity and cumbersomeness and supporting quick learning. The total score reached a level between good and excellent when compared with published systems used for similar purposes.

The dedicated generative artificial intelligence system provides a structured approach to customised patient scheduling in hospitals and clinics. It imports scheduling data, accepts natural-language requirements, uses a deep neural network parser and automatically generates a genetic algorithm to solve the resulting problem. In the day surgery case, it achieved high parsing performance, rapid solution times and scheduling performance above current practices. The system also has limitations, including black-box concerns, possible multi-objective requirements and challenges in scaling to larger hospitals or additional healthcare resources.

Source: Health Informatics Journal