Identifying primary-care features associated with complex mental health difficulties.

Area of Science:

• Health Informatics
• Psychiatry
• Data Science

Background:

• Coded prevalence of complex mental health difficulties in electronic health records (EHRs) is significantly lower than population estimates.
• This discrepancy suggests underdiagnosis or poor coding of conditions like personality disorder and dysthymia.

Purpose of the Study:

• To identify features within primary care records that could improve the recognition of complex mental health difficulties.
• To explore the utility of information theoretic approaches for feature generation in EHRs.

Main Methods:

• Analysis of the Connected Bradford primary care database (approx. 1.15M citizens).
• Generation of over 500,000 features using multi-level patient attributes, including psychiatric, social, and medical problems, prescription patterns, service use, and temporal attendance stability.
• Application of scaled mutual information to assess the relationship between features and diagnostic codes.

Main Results:

• Identified 3,040 records (0.3%) meeting criteria for complex mental health difficulties, far below the expected 3-5% prevalence.
• The count of unique psychiatric diagnoses was the most informative feature.
• Other significant features included antipsychotic prescription status, non-attendance entropy, and counts of self-harm and substance misuse behaviors, though positive predictive value was often low.

Conclusions:

• Complex mental health conditions are inadequately coded in EHRs.
• Information theoretic methods can generate novel, potentially useful features from EHR data.
• While not sufficient for diagnosis, these features can serve as valuable prompts for further clinical assessment.