Mihael Drinovac1, Helga Wagner1, Niruj Agrawal2, Hannah R Cock3, Alex J Mitchell4, Tim J von Oertzen5. 1. Institute of Applied Statistics, Johannes Kepler University, Linz, Austria. 2. Department of Neuropsychiatry, St George's Hospital, London, UK; Epilepsy Group, Atkinson Morley Regional Neurosciences Centre, St George's Hospital, London, UK; St George's University of London, London, UK. 3. Epilepsy Group, Atkinson Morley Regional Neurosciences Centre, St George's Hospital, London, UK; St George's University of London, London, UK. 4. Department of Cancer Studies and Molecular Medicine, University of Leicester, Leicester, UK; Department of Psycho-oncology, Leicestershire Partnership NHS Trust, Leicester, UK. 5. St George's University of London, London, UK; Department of Neurology, Wagner-Jauregg Neuroscience Centre, Kepler University Hospital, Linz, Austria. Electronic address: neurologiesekr.wj@gespag.at.
Abstract
OBJECTIVE: Depression is common but frequently underdiagnosed in people with epilepsy. Screening tools help to identify depression in an outpatient setting. We have published validation of the NDDI-E and Emotional Thermometers (ET) as screening tools for depression (Rampling et al., 2012). In the current study, we describe a model of an optimized screening tool with higher accuracy. METHODS: Data from 250 consecutive patients in a busy UK outpatient epilepsy clinic were prospectively collected. Logistic regression models and recursive partitioning techniques (classification trees, random forests) were applied to identify an optimal subset from 13 items (NDDI-E and ET) and provide a framework for the prediction of class membership probabilities for the DSM-IV-based depression classification. RESULTS: Both logistic regression models and classification trees (random forests) suggested the same choice of items for classification (NDDI-E item 4, NDDI-E item 5, ET-Distress, ET-Anxiety, ET-Depression). The most useful regression model includes all 5 mentioned variables and outperforms the NDDI-E as well as the ET with respect to AUC (NDDI-E: 0.903; ET7: 0.889; logistic regression: 0.943). A model developed using random forests, grown by restricting the possible splitting of variables to these 5 items using only subsets of the original data for single classification, performed similarly (AUC: 0.949). CONCLUSIONS: For the first time, we have created a model of a screening tool for depression containing both verbal and visual analog scales, with characteristics supporting that this will be more precise than previous tools. Collection of a new data sample to assess out-of-sample performance is necessary for confirmation of the predictive performance.
OBJECTIVE:Depression is common but frequently underdiagnosed in people with epilepsy. Screening tools help to identify depression in an outpatient setting. We have published validation of the NDDI-E and Emotional Thermometers (ET) as screening tools for depression (Rampling et al., 2012). In the current study, we describe a model of an optimized screening tool with higher accuracy. METHODS: Data from 250 consecutive patients in a busy UK outpatientepilepsy clinic were prospectively collected. Logistic regression models and recursive partitioning techniques (classification trees, random forests) were applied to identify an optimal subset from 13 items (NDDI-E and ET) and provide a framework for the prediction of class membership probabilities for the DSM-IV-based depression classification. RESULTS: Both logistic regression models and classification trees (random forests) suggested the same choice of items for classification (NDDI-E item 4, NDDI-E item 5, ET-Distress, ET-Anxiety, ET-Depression). The most useful regression model includes all 5 mentioned variables and outperforms the NDDI-E as well as the ET with respect to AUC (NDDI-E: 0.903; ET7: 0.889; logistic regression: 0.943). A model developed using random forests, grown by restricting the possible splitting of variables to these 5 items using only subsets of the original data for single classification, performed similarly (AUC: 0.949). CONCLUSIONS: For the first time, we have created a model of a screening tool for depression containing both verbal and visual analog scales, with characteristics supporting that this will be more precise than previous tools. Collection of a new data sample to assess out-of-sample performance is necessary for confirmation of the predictive performance.
Authors: Alejandro L Vázquez; Melanie M Domenech Rodríguez; Tyson S Barrett; Sarah Schwartz; Nancy G Amador Buenabad; Marycarmen N Bustos Gamiño; María de Lourdes Gutiérrez López; Jorge A Villatoro Velázquez Journal: Prev Sci Date: 2020-02