Seetal Dodd1, Michael Berk2, Katarina Kelin3, Qianyi Zhang4, Elias Eriksson5, Walter Deberdt6, J Craig Nelson7. 1. Department of Psychiatry, University of Melbourne, Parkville, VIC, Australia; IMPACT SRC (Innovation in Mental and Physical Health and Clinical Treatment), School of Medicine, Deakin University, Geelong, VIC, Australia. Electronic address: seetald@barwonhealth.org.au. 2. Department of Psychiatry, University of Melbourne, Parkville, VIC, Australia; IMPACT SRC (Innovation in Mental and Physical Health and Clinical Treatment), School of Medicine, Deakin University, Geelong, VIC, Australia; Florey Institute for Neuroscience and Mental Health, University of Melbourne, Parkville, VIC, Australia; ORYGEN Research Centre, Parkville, VIC, Australia. 3. Eli Lilly Australia Pty Limited, West Ryde, NSW, Australia. 4. Eli Lilly and Company, Indianapolis, IN, USA. 5. Department of Pharmacology and the Institute of Physiology and Neuroscience, Sahlgrenska, Academy, Goteborg University, Goteborg, Sweden. 6. Eli Lilly and Company, Brussels, Belgium. 7. University of California, San Francisco, CA, USA.
Abstract
BACKGROUND: Randomised, placebo-controlled trials of treatments for depression typically collect outcomes data but traditionally only analyse data to demonstrate efficacy and safety. Additional post-hoc statistical techniques may reveal important insights about treatment variables useful when considering inter-individual differences amongst depressed patients. This paper aims to examine the Gradient Boosted Model (GBM), a statistical technique that uses regression tree analyses and can be applied to clinical trial data to identify and measure variables that may influence treatment outcomes. METHODS: GBM was applied to pooled data from 12 randomised clinical trials of 4987 participants experiencing an acute depressive episode who were treated with duloxetine, an SSRI or placebo to predict treatment remission. Additional analyses were conducted on the same dataset using the logistic regression model for comparison between these two methods. RESULTS: With GBM, there were noticeable differences between treatments when identifying which and to what extent variables were associated with remission. A single logistic regression only revealed a decreasing or increasing relationship between predictors and remission while GBM was able to reveal a complex relationship between predictors and remission. LIMITATIONS: These analyses were conducted post-hoc utilising clinical trials databases. The criteria for constructing the analyses data were based on the characteristics of the clinical trials. CONCLUSIONS: GBM can be used to identify and quantify patient variables that predict remission with specific treatments and has greater flexibility than the logistic regression model. GBM may provide new insights into inter-individual differences in treatment response that may be useful for selecting individualised treatments. TRIAL REGISTRATION: IMPACT clinical trial number 3327; IMPACT clinical trial number 4091; IMPACT clinical trial number 4689; IMPACT clinical trial number 4298; NCT00071695; NCT00062673; NCT00036335; NCT00067912; NCT00073411; NCT00489775; NCT00536471; NCT00666757 (note that trials with IMPACT numbers predate mandatory clinical trial registration requirements).
BACKGROUND: Randomised, placebo-controlled trials of treatments for depression typically collect outcomes data but traditionally only analyse data to demonstrate efficacy and safety. Additional post-hoc statistical techniques may reveal important insights about treatment variables useful when considering inter-individual differences amongst depressedpatients. This paper aims to examine the Gradient Boosted Model (GBM), a statistical technique that uses regression tree analyses and can be applied to clinical trial data to identify and measure variables that may influence treatment outcomes. METHODS: GBM was applied to pooled data from 12 randomised clinical trials of 4987 participants experiencing an acute depressive episode who were treated with duloxetine, an SSRI or placebo to predict treatment remission. Additional analyses were conducted on the same dataset using the logistic regression model for comparison between these two methods. RESULTS: With GBM, there were noticeable differences between treatments when identifying which and to what extent variables were associated with remission. A single logistic regression only revealed a decreasing or increasing relationship between predictors and remission while GBM was able to reveal a complex relationship between predictors and remission. LIMITATIONS: These analyses were conducted post-hoc utilising clinical trials databases. The criteria for constructing the analyses data were based on the characteristics of the clinical trials. CONCLUSIONS: GBM can be used to identify and quantify patient variables that predict remission with specific treatments and has greater flexibility than the logistic regression model. GBM may provide new insights into inter-individual differences in treatment response that may be useful for selecting individualised treatments. TRIAL REGISTRATION: IMPACT clinical trial number 3327; IMPACT clinical trial number 4091; IMPACT clinical trial number 4689; IMPACT clinical trial number 4298; NCT00071695; NCT00062673; NCT00036335; NCT00067912; NCT00073411; NCT00489775; NCT00536471; NCT00666757 (note that trials with IMPACT numbers predate mandatory clinical trial registration requirements).
Authors: Malcolm P Forbes; Adrienne O'Neil; Melissa Lane; Bruno Agustini; Nick Myles; Michael Berk Journal: Drugs Aging Date: 2021-04-29 Impact factor: 3.923
Authors: J E J Buckman; R Saunders; J Stott; L-L Arundell; C O'Driscoll; M R Davies; T C Eley; S D Hollon; T Kendrick; G Ambler; Z D Cohen; E Watkins; S Gilbody; N Wiles; D Kessler; D Richards; S Brabyn; E Littlewood; R J DeRubeis; G Lewis; S Pilling Journal: Epidemiol Psychiatr Sci Date: 2021-06-04 Impact factor: 6.892
Authors: Joshua E J Buckman; Rob Saunders; Zachary D Cohen; Phoebe Barnett; Katherine Clarke; Gareth Ambler; Robert J DeRubeis; Simon Gilbody; Steven D Hollon; Tony Kendrick; Edward Watkins; Nicola Wiles; David Kessler; David Richards; Deborah Sharp; Sally Brabyn; Elizabeth Littlewood; Chris Salisbury; Ian R White; Glyn Lewis; Stephen Pilling Journal: Psychol Med Date: 2021-04-14 Impact factor: 7.723