Quantitative Assessment of the Exposure Risk and Targeted Protection of Sensitive Information related to Mental Health Disorders in China's Electronic Medical Records

Background: Mental health issues affect populations worldwide, with depression, schizophrenia, and dementia being particularly prevalent in China, where the China Mental Health Survey (CMHS) reported a 7.4% lifetime prevalence of mood disorders. Mental disorders have become a leading cause of disability. While the widespread adoption of electronic medical records (EMRs) has significantly improved healthcare efficiency and resource allocation, the sensitivity of medical data poses serious privacy breach risks. Many patients withhold medical information due to data security concerns, increasing the risk of treatment discontinuation. Currently, the lack of unified management policies and technical standards for electronic health records (EHRs) has led to frequent unauthorized access, leaks, and illegal trading of data, exacerbating doctor-patient conflicts and societal stigma against individuals with mental illness. Methods: This study developed a novel personally identifiable information (PII) desensitization protocol (EPPDI) to mitigate privacy risks through comprehensive database scanning (as opposed to traditional field-specific desensitization). The protocol incorporates the following technological innovations: (1) Expansion of a lexicon of 20 mental health-related keywords using a Word2Vec vector space model; (2) Application of regular expressions to replace sensitive information and its surrounding 10 characters with asterisks (*). Results: Among 1,235,651 patients (8,016,263 records), the EPPDI protocol achieved 97.60% precision and 95.40% recall, with a privacy protection efficacy rate of 97.85%. Diagnosis records (31.84%) and medication data (45.41%) were identified as primary leakage sources. Regional disparities were notable, with Beijing showing a PD identification rate of 21.06%, far exceeding Qinghai's 1.66%. Regarding data utility preservation, among 2,000 pieces of PDI patient information, 48 (2.4% false positive rate) contained no sensitive content, while 92 (4.6% false negative rate) pieces of non-PDI patient information included sensitive data among 2,000 pieces of non-PDI patient information. Conclusion: The EPPDI protocol addresses challenges such as ambiguity in Chinese terminology and adaptation to unstructured narratives, providing a technical framework for implementing China's Personal Information Protection Law in mental health. Future efforts should focus on balancing privacy protection with research needs through dynamic, tiered desensitization approaches.