Identifying primary-care features associated with complex mental health difficulties

Abstract

Aim: The coded prevalence of complex mental health difficulties in electronic health records, such as personality disorder and dysthymia,is much lower than expected from population surveys. We aimed to identify features in primary care records that might be useful in promoting greater recognition of complex mental health difficulties.

Methods and findings: We analysed Connected Bradford, an anonymised primary care database of approximately 1.15M citizens. We used multiple approaches to generate a large set of features representing multi-level collections of patient attributes across time and dimensions of healthcare. Feature sets included antecedent and concurrent problems (psychiatric, social and medical), patterns of prescription and service use and temporal stability of attendance. These were tested individually and in combination. We analysed the relationship between features and diagnostic codes using scaled mutual information. We identified 3,040 records satisfying our definition of complex mental health difficulties. This was 0.3% of the population compared to an expected prevalence of 3-5%. We generated >500,000 features. The most informative feature was count of unique psychiatric diagnoses. Other features were identified, including binary features (e.g., presence or absence of prescription for antipsychotic medication), continuous features (e.g., entropy of non-attendance) and counts of features (e.g., concerning behaviours such as self-harm & substance misuse). Several of these showed odds ratios >=5 or <=0.2 but low positive predictive value. We suggest this is due to the large number of "cases" being uncoded and, thus appearing as "controls".

Conclusion: Complex mental health difficulties are poorly coded. We demonstrated the feasibility of using information theoretic approaches to develop a large set of novel features in electronic health records. While these are currently insufficient for diagnosis, several can act as prompts to consider further diagnostic assessment.

Fig 1. Visualisation of groups defined by information within electronic healthcare records.

Fig 2. Examples of the distributions of values for some component features.

Fig 3. Examples of the distributions of values for some feature families.

Fig 4. Scaled mutual information for all informative feature sets in rank order.

Rank is presented in log₁₀ to illustrate how tightly packed the scaled mutual information scores were across orders of magnitude of rank (illustrated by the straight-line fit). A single, outstanding feature showed a scaled mutual information value greater than 8.2%: the count of psychological disorders.

Fig 5. Scaled mutual information for all informative feature sets in rank order.

Rank is presented in log₁₀. A) Rank of component feature sets. B) Rank of family feature sets. C) Rank of feature-family combinations.