West China Hospital of Sichuan University runs an AI voice follow-up system that calls discharged patients automatically. Live since April 2024, the system had by October 2024 reached 99% follow-up coverage of discharged patients; 88% of discharged patients answered the follow-up call, and 83% of AI voice calls were completed successfully. In the third quarter of 2024 the hospital discharged more than 60,000 patients and had more than 130,000 follow-up call tasks. The voice robot can dial 40 calls at the same time, which raised call-execution efficiency about 40-fold and removed 106,000 manual calls from nurses' workload.
Follow-up pressure was heavy. The hospital had more than 20,000 discharges a month in 2023, and the Sichuan provincial patient-experience plan for 2023-2025 requires tertiary public hospitals to reach an 80% discharge follow-up rate by the end of 2024 and 90% by the end of 2025. Before the system, a manual follow-up call took about 10 to 15 minutes per patient because of patients' accents and communication difficulties. With scripts tailored to each stage that ask about body temperature, wounds, diet and sleep, an AI voice follow-up takes about 2 minutes per patient, roughly 85% less time.
The project started in December 2023 and built a patient profile base, a knowledge base and a basic platform. After a pilot the system was deployed hospital-wide. Two weeks after launch the management department collected feedback from clinical departments, and a shared group linking management, clinical departments, the information center and the vendor handles questions. With the extra capacity the hospital changed its follow-up mechanism from one call to two, on day 7 and day 30 after discharge.
The system combines acoustic front-end processing, speech recognition, speech synthesis, speech content extraction and analysis, content understanding and summarization, multi-turn dialogue management, natural language processing, big-data analysis and cloud computing. It connects to the hospital information system: admission and discharge data are passed through an interface, the system matches a follow-up path automatically from the patient's discharge department and surgery, and calls are placed at each task node without adding operations for medical staff. Results are analysed by department, disease and time period, including the reasons calls fail, and staff track by hand the patients whose calls failed because of wrong numbers, unanswered calls or dropped calls.
The system covers 18 departments and 81 inpatient nursing units, including laboratory medicine, the admission service center and the chronic disease management center. Follow-up answers also flag patients with research value, who are recommended to the research platform for enrolment after a doctor's assessment, and patients who qualify for chronic disease management are recommended to that platform.
Next, the hospital intends to build an AI follow-up system with deep-learning ability that extends follow-up to whole-course patient management and coordination across its medical alliance. It would keep a knowledge graph updated from newly published diagnosis and treatment guidelines, expert consensus on post-discharge follow-up and a systematic disease follow-up database, build a profile for each patient from their medical record and, combined with disease-specific knowledge bases, generate individual follow-up schedules and health education content. The hospital also intends to integrate the outbound follow-up calls with a large language model, so that replies are generated from the patient's meaning, the standard scripts and the conversation context, and when an answer is not understood, the model reasons from the context and the dialogue flow chart to decide the next step and how to respond.