Investigation of negative symptoms in schizophrenia with a machine learning text-mining approach

Background
Negative symptoms account for the greatest burden of illness among individuals with schizophrenia. These symptoms are an increasingly important target for therapy, especially since their presence predicts poor long-term clinical outcomes. However, development of practicable methods for their assessment has been difficult. In this study, we present a novel support vector machine learning method to see whether we could identify the presence of negative symptoms in electronic health records.

Methods
We used routinely collected clinical data from the Biomedical Research Council Case Register, South London and Maudsley NHS Trust. Data were obtained from the case records of 7678 adults with schizophrenia receiving care in 2011, of whom 1590 were inpatients. A training dataset of around 200 case records from this sample was analysed with the General Architecture for Text Engineering Machine Learning software package to develop a text-mining tool that was subsequently used to estimate the prevalence of negative symptoms in the whole sample. Multivariable logistic and multiple linear regression analyses were done to investigate the association of negative symptomatology with age, sex, relationship status, impairment of activities of daily living, and (for inpatients) length of hospital stay.

Findings
4269 patients (55·7%) had at least one negative symptom documented. Negative symptoms were particularly associated with patients who were aged 20—29 years (all other age groups odds ratio [OR] and upper 95% CI limit <1·0, p<0·001), male (OR 1·29, 95% CI 1·17—1·44; p<0·001), and not in a relationship (1·31, 1·11—1·56; p=0·002). They were also associated with impairment of activities of daily living (1·35, 1·21—1·52; p<0·001) and increased likelihood of hospital admission (1·24, 1·10—1·39; p<0.001). Among inpatients, emotional withdrawal (β=30·0, 95% CI 15·6—44·4; p<0·001) and apathy (27·4, 1·8—53·1; p=0·036) were particularly associated with increased length of stay in hospital.

Interpretation
Using a machine learning approach, we were able to identify the presence of negative symptoms in electronic health records. The data suggest that negative symptoms are evident in most patients with schizophrenia and are associated with poor clinical outcomes. These findings highlight the need for the development of new treatments that can alleviate negative symptoms. Furthermore, the increasing use of electronic health records highlights an opportunity to adopt support vector machine learning text-mining approaches to obtain data for research and clinical decision support in other areas of medicine.

Funding
UK Medical Research Council, National Institute for Health Research, Roche.