From Medical Records to Data Utilization. Development of an artificial intelligence process to structure unstructured medical data in thyroid pathology

This thesis aims to develop artificial intelligence (AI) methods to extract, structure, and analyze unstructured data from electronic health records (EHR) in thyroid pathology. These health data warehouses are crucial for both research and clinical care, facilitating the management of complex information derived from various medical reports.In the first phase, a convolutional neural network (CNN) algorithm was trained on a dataset of 6,000 patients to identify those with thyroid pathologies. The algorithm's performance was evaluated using key metrics such as accuracy (0.95), recall (0.83), precision (0.85), specificity (0.97), and F1 score (0.84), demonstrating its high robustness in identifying relevant patients.The second phase involved automating the extraction of data from various medical reports, including consultations, ultrasounds, anesthesia records, surgical reports, and histopathology. An AI model was tested on a cohort of 1,500 patients to extract 42 clinical variables. The overall accuracy of the tool was 90.89%. Compared to human experts (97.95%) and non-experts (95.45%), the algorithm exhibited similar performance for most variables. However, its performance was less accurate for some more complex variables requiring clinical judgment, underscoring the need for human oversight in more nuanced tasks.Concurrently, an interactive statistical analysis tool, StatiCAL, was developed to enable real-time descriptive, univariate, and multivariate analyses on the generated dataset. Accessible via the Centre Antoine Lacassagne's intranet, this tool allows users to conduct analyses without programming skills, making statistical analysis more accessible while ensuring robust and reliable results.Additionally, a retrospective study was conducted on 3,334 patients to identify predictive factors for the loss of neuromonitoring signal (LOS) during thyroid surgery and its impact on recurrent laryngeal nerve palsy (RLNP). The findings revealed that male gender (OR = 2.64) and total thyroidectomy (OR = 1.76) were independent risk factors for LOS. However, these factors did not significantly affect the occurrence of RLNP, as only 41% of patients who experienced LOS developed RLNP.In conclusion, this thesis demonstrates that AI can significantly enhance the extraction and exploitation of medical data, optimizing data management for both clinical research and medical practice.