Introduction
Antidepressant monotherapy, such as selective serotonin reuptake inhibitors (SSRIs), represents the most common first-step treatment for depression, accounting for 87% of delivered treatment modalities (Olfson, Blanco, & Marcus, Reference Olfson, Blanco and Marcus2016). Unfortunately, only 28–33% of individuals receiving SSRI monotherapy achieve successful remission (Trivedi et al., Reference Trivedi, Rush, Wisniewski, Nierenberg, Warden and Ritz2006), indicating that a majority of patients are still symptomatic and require further treatment.
There are limited guidelines to determine which treatment strategy would be optimal for a given individual who does not successfully respond to an initial course of SSRI. Unfortunately, available guidelines are derived from nomothetic evidence collected across group averages, which obfuscates individual differences that are better addressed by more idiographic approaches. Common principles of these nomothetic guidelines, as embodied in Bennabi et al. (Reference Bennabi, Charpeaud, Yrondi, Genty, Destouches, Lancrenon and Dorey2019) and Gelenberg et al. (Reference Gelenberg, Freeman, Markowitz, Rosenbaum, Thase, Trivedi and Van Rhoads2010), have included switching treatment after no response to initial treatment and augmenting treatment after partial response to initial treatment. Such recommendations are often general in nature and are applied in combination with clinical intuition. In practice, the sequencing of treatment selection frequently results in a protracted process of trial and error in identifying the most appropriate treatment for an individual patient. As a consequence, burden is exacerbated both in terms of duration of illness and number of provider visits (Cusin & Peyda, Reference Cusin, Peyda, Shapero, Mischoulon and Cusin2019; Kazdin & Blase, Reference Kazdin and Blase2011).
Recently, mental health research has increasingly embraced precision medicine as a framework for personalizing treatment using more computationally rigorous procedures (Bernardini et al., Reference Bernardini, Attademo, Cleary, Luther, Shim, Quartesan and Compton2017). A precision medicine approach leveraging machine learning may facilitate the effort to reduce the amount of time required to identify the optimal second-step treatment strategy. Commonly collected information from a routine psychiatric assessment visit (e.g. demographics, questionnaires, diagnostic information, prior treatment history, etc.) could be provided to machine learning algorithms to predict whether a specific second-step treatment strategy (i.e. augmentation or switching) would improve the probability of remission. Such a machine learning framework fosters a more idiographic approach to second-step treatment selection relative to standard nomothetic guidelines.
Prior efforts to apply machine learning as a precision medicine approach to depression treatment have illustrated how disease burden can be mitigated. One study validated a machine learning algorithm for initial treatment remission for depression that outperformed the baseline remission rate by 11% (Mehltretter et al., Reference Mehltretter, Fratila, Benrimoh, Kapelner, Perlman, Snook and Turecki2020). Assuming five visits per treatment for each patient, it was estimated that this improvement in prediction could result in approximately 1772 fewer doctor's visits and decreased duration of disease burden.
The current study introduces a computationally principled framework for improving predictions about patient treatment remission using ensemble machine learning with the super learner algorithm. A principal advantage of ensemble learners is that they leverage several individual algorithms to obtain better performance than any of the constituent learners alone. Specifically, ensemble machine learning using a super learner approach involves consolidating several individual machine learning algorithms into an ensemble algorithm that represents a weighted average of all individual learners (e.g. random forest, elastic net regression, neural network, etc.). Thus, rather than performing predictive modeling with several individual machine learning algorithms and selecting the best performing algorithm, as is common practice, the ensemble approach capitalizes on the individual strengths of several disparate learning algorithms by including them all in the final model. The ensemble machine learning algorithms will be applied to the largest, multi-center clinical trial ever conducted on treatments for depression: the STAR*D study. To date, machine learning approaches to depression treatment have been primarily used for first-step interventions of standard care (e.g. SSRI monotherapy) (Chekroud et al., Reference Chekroud, Zotti, Shehzad, Gueorguieva, Johnson, Trivedi and Corlett2016; Mehltretter et al., Reference Mehltretter, Fratila, Benrimoh, Kapelner, Perlman, Snook and Turecki2020). This project represents the first study to employ an ensemble machine learning approach to improve predictions of treatment response to one of seven second-step interventions for depression. Precision medicine may facilitate more expeditious provision of care that is personalized to specific depressed individuals. Specifically, this study has the potential to inform efforts to develop clinical decision-support tools that could improve patient outcomes.
Methods
The current study is a secondary data analysis of the STAR*D study, which was a multicenter longitudinal NIMH-sponsored study. STAR*D was designed to determine the short- and long-term effects of different sequences of medication and/or psychotherapy for the treatment of unipolar depression that has not responded adequately to an initial standard antidepressant trial. In particular, STAR*D compared the effectiveness of different sets of treatment options: (1) augmenting the first antidepressant with another medication or psychotherapy, (2) switching to another antidepressant or psychotherapy.
Trial design
At Level 1, patients initially received citalopram (CIT) for a minimum of 8 weeks. Patients who did not remit during Level 1 were offered Level 2 treatment, which represented two overall treatment strategies: (1) Medication or Psychotherapy Switch – switching from CIT to another antidepressant medication or Cognitive Psychotherapy (CT), and (2) Medication or Psychotherapy Augmentation – augmenting CIT with a second medication or CT. For those who switch treatments at Level 2, sertraline (SER) (a second SSRI), venlafaxine (VEN) (an antidepressant with both noradrenergic and serotonergic effects), bupropion (BUP) (an antidepressant with both noradrenergic and dopaminergic effects), or CT were available. Likewise, within the Medication or Psychotherapy Augmentation strategy, the three treatments for augmenting CIT were BUP, buspirone (BUS) (an antianxiety medication), or CT.
Study inclusion and exclusion criteria
Inclusion criteria permitted outpatients who were 18–75 years of age and had a nonpsychotic major depressive episode determined by a baseline HAM-D score ⩾14. They were eligible if their clinicians determined that outpatient treatment with an antidepressant medication was both safe and indicated. Patients who were pregnant or breast-feeding and those with a primary diagnosis of bipolar, psychotic, obsessive-compulsive, or eating disorders were excluded from the study, substance dependence (only if it required inpatient detoxification) or a clear history of nonresponse or intolerance to any protocol antidepressant in the first two treatment steps were also exclusionary. To ensure broad and inclusive eligibility criteria, co-morbidity with other Axis I anxiety and emotional disorders was permitted. Diagnostic criteria were assessed with a clinician administered checklist measuring Axis I symptoms of the Diagnostic and Statistics Manual, fourth edition revised (Trivedi et al., Reference Trivedi, Rush, Wisniewski, Nierenberg, Warden and Ritz2006).
Participants
In Level 1 of STAR*D, 4041 depressed patients enrolled, and, after exclusion criteria were applied and study discontinuation was considered, 2876 patients were included in the Level 1 results. In Level 2, 1439 patients were randomized to one of the seven treatment options deemed acceptable to each participant. Patients were 63.7% female, and the mean age was 40.8 (s.d. = 13). Regarding self-reported ethnicity, 75.8% of patients were white, 17.6% were African American, and 6.6% identified as other.
Measures
In the current analysis, the predictor variables were all assessed prior to initiation of Level 2 treatments. Specifically, demographic predictors were collected at the very onset of the STAR*D study design, and clinical assessment variables were collected subsequent to Level 1 treatment, but prior to Level 2 treatment. The outcome variable (i.e. treatment remission) was estimated post-treatment for each Level 2 intervention.
Predictor variables
Demographic and psychosocial variables
Includes age, sex, race, marital status, education in years, occupational employment status, comorbid disorders, age of depression onset, and length of illness.
Clinical assessment variables
Includes the Quick Inventory of Depressive Symptomatology (QIDS; Rush et al., Reference Rush, Trivedi, Ibrahim, Carmody, Arnow, Klein and Thase2003), Hamilton Rating Scale for Depression (HAMD; Hamilton, Reference Hamilton1960), Short-Form Health Survey (SF-12; Ware, Kosinski, and Keller, Reference Ware, Kosinski and Keller1996), Work and Social Adjustment Scale (WSAS; Mundt, Marks, Shear, and Greist, Reference Mundt, Marks, Shear and Greist2002), Work Productivity and Activity Impairment Questionnaire (WPAI; Reilly, Zbrozek, and Dukes, Reference Reilly, Zbrozek and Dukes1993), Quality of Life Enjoyment and Satisfaction Questionnaire (Q-LES-Q; Endicott, Nee, Harrison, and Blumenthal, Reference Endicott, Nee, Harrison and Blumenthal1993), and the first 100 items of the Psychiatric Diagnostic Symptom Questionnaire (PDSQ; Zimmerman and Mattia, Reference Zimmerman and Mattia2001).
Outcome variables
Treatment remission
Consistent with the original study (Trivedi et al., Reference Trivedi, Rush, Wisniewski, Nierenberg, Warden and Ritz2006), treatment remission is defined as a post-treatment score of ⩽ 7 on the total HAM-D scale, treated as a binary, dummy-coded outcome variable (i.e. remitted = 1, non-remitted = 0).
Data analysis
To predict treatment remission, machine learning algorithms were evaluated using all the available predictor variables listed above (total = 115). Listwise deletion was applied prior to final analyses, as the machine learning strategy currently used requires no missing data. Specifically, a super learner approach was adopted to optimize prediction performance. Super learning involves consolidating several individual machine learning algorithms into a stacked ensemble algorithm, representing a weighted average of all individual learners (Polley, LeDell, Kennedy, & Laan, Reference Polley, LeDell, Kennedy and Laan2021; Van der Laan, Polley, & Hubbard, Reference Van der Laan, Polley and Hubbard2007; Van der Laan & Rose, Reference Van der Laan and Rose2011). The final aggregate model is at least as accurate as the best individual algorithm included in the ensemble (Boehmke & Greenwell, Reference Boehmke and Greenwell2019). The ensemble model combined several well-established algorithms designed for categorical outcomes that cover a variety of analytical assumptions: elastic net binomial regression, ranger random forest, support vector machine, neural network, extreme gradient boosting, Bayesian generalized linear regression, and the overall mean (Boehmke & Greenwell, Reference Boehmke and Greenwell2019). For a brief description of each procedure please refer to online Supplementary Appendix I in the Supplementary Materials. Furthermore, a t test variable filtering procedure was employed for only the training sample folds to identify the top ten predictors for each treatment condition. In brief, t test variable filtering examines the relationship between the binary outcome variable and the predictor variables in the training folds by preforming t tests with the binary outcome variable of the machine learning model now being used as the independent variable. The rank order of the t test p values are determined across all of the machine learning predictor variables, and the top ten most significant predictor variables were retained. The reasons for variable filtering were twofold. First, variable filtering can mitigate the influence of excessive irrelevant and non-informative predictors, which can undermine model performance and increase error for several algorithms (Kuhn & Johnson, Reference Kuhn and Johnson2019). Second, deriving models with fewer predictors can facilitate future implementation.
When predicting categorical classification outcomes, it is important to consider the presence of potential class imbalances in the outcome variable. Large class imbalances can bias machine learning results toward favoring prediction of the more frequent outcome. Class imbalances (i.e. minority class representing less than 25%) were observed for the CITCT and CT conditions. To address this, we implemented the Synthetic Minority Over-sampling Technique (SMOTE), which combines over-sampling of the minority class and under-sampling of the majority class to promote a more balanced class structure for the outcome variable (Chawla, Bowyer, Hall, & Kegelmeyer, Reference Chawla, Bowyer, Hall and Kegelmeyer2002).
Machine learning models were examined using nested cross-validation, which partitions the sample into outer loop and an inner loop subsets. For the outer loop, ten folds were specified, of which nine are used in the training process and final predictions are made in the remaining subset. This process is repeated for each of the remaining subsets, and results are averaged to produce a single estimate. The inner loop also contains ten folds. The inner loop is used to build the ensemble estimator and the weighting of the constituent individual learners, whereas the outer loop test set is used to provide a final estimation of the performance of the ensemble model. Nested cross-validation ensures that the final testing folds are not contaminated as a consequence of data leakage. Variable filtering and SMOTE procedures were applied to only the training folds. Separate ensemble machine learning models were estimated for each of the seven treatment conditions.
Performance metrics included area under the receiver operator curve (AUC, a measure of model discrimination), sensitivity, specificity, positive predictive value (PPV, i.e. the probability that a patient predicted to achieve remission actually does so), negative predictive value (NPV i.e. the probability that a patient predicted to not achieve remission actually does not), and accuracy (the proportion of total outcomes – remitted v. not remitted--classified correctly). PPV and NPV were determined using the sensitivity and specificity values empirically estimated from the final model performance for each model. The class probability threshold was 0.50 or greater for determining predicted remission status. Analyses were performed using the following R packages: nestedcv (Lewis, Spiliopoulou, & Goldmann, Reference Lewis, Spiliopoulou and Goldmann2022), SuperLearner (Polley et al., Reference Polley, LeDell, Kennedy and Laan2021), and Caret (Kuhn, Reference Kuhn2022).
Results
Model performance with full predictor Set
Results of the final ensemble models with nested cross-validation revealed differential classification performance across treatment type (Table 1). These results are visually depicted in Fig. 1. Discrimination was significantly better-than-chance for CT (AUC = 0.82, 95% CI 0.64–0.99) and BUP (AUC = 0.66, 95% CI 0.58–0.74). However, sensitivity was poor for the BUP model (5%) and moderate (33%) for the CT model. The remaining super learner models yielded AUCs that did not outperform chance: CITBUP (AUC = 0.52, 95% CI 0.45–0.59), CITBUS (AUC = 0.55, 95% CI 0.44–0.63), CITCT (AUC = 0.51, 95% CI 0.36–0.66), SER (AUC = 0.57, 95% CI 0.48–0.66), and VEN (AUC = 0.57, 95% CI 0.49–0.65).
Predictive performance with full predictor set
Note: *** = p < 0.001; AUC, area under the curve; CI, confidence interval; BUP, Bupropion; CITBUP, Citalopram + Bupropion; CITBUS, Citalopram + Buspirone; CITCT, Citalopram + Cognitive Therapy; CT, Cognitive Therapy; SER, Sertraline; VEN, Venlafaxine.
ROC curve for models with full predictor set.
Model performance with filtered predictor set
As a consequence of the t test filter, each super learner model was trained on the top ten predictors most associated with remission. In Table 2, the top ten predictors are displayed for each intervention. Of note, the variables that were most commonly predictive across the treatments include (1) work and social related impairment (WSAS), which was predictive for four treatments (i.e. BUP, CITBUS, SER, and VEN); (2) quality of life (QLESQ), which was predictive for four treatments (i.e. BUP, CITBUS, CITCT, and VEN); depression symptomatology as measured by the QIDS (QVTOT), which was predictive for four treatments (i.e. BUP, CITBUS, CT, and SER); and overall depression severity as measured by the HDRS (HDTOT), which was predictive for three treatments (i.e. CITBUP, SER, and VEN).
Top ten filtered variables
BUP, bupropion; CITBUP, citalopram + Bupropion; CITBUS, citalopram + cuspirone; CITCT, citalopram + cognitive therapy; CT, cognitive therapy; SER, sertraline; VEN, venlafaxine.
Variables are filtered based on the ranking of the smallest t test p values. Unless otherwise specified (i.e. HAMD, QIDS, WSAS, QLESQ, QPAI, SF-12, sex, and years of schooling), all variables denote individual items from the PDSQ.
Note: Exponentiated coefficients (i.e. odds ratios) are presented from the Bayesian generalized linear model algorithm to facilitate interpretation. * = p < 0.05, ** = p < 0.01, *** = p < 0.001.
For models with the top ten filtered variables, results revealed classification performance as significantly better than chance for five of the second-step treatments (Table 3). These results are visually depicted in Fig. 2. The best predictive performance was still associated with the CT (AUC = 0.72, 95% CI 0.504–0.94) and BUP (AUC = 0.70, 95% CI 0.62–0.79) treatments. Furthermore, statistically significant prediction was obtained for CITCT (AUC = 0.65, 95% CI 0.501–0.81), VEN (AUC = 0.60, 95% CI 0.51–0.68), and SER (AUC = 0.59, 95% CI 0.51–0.67). The remaining two super learner models were associated with AUC performance that did not outperform chance: CITBUP (AUC = 0.55, 95% CI 0.47–0.62) and CITBUS (AUC = 0.53, 95% CI 0.46–0.61).
Predictive performance with top ten predictors
Note: * = p < 0.05; *** = p < 0.001; AUC, area under the curve; CI, confidence interval; BUP, bupropion; CITBUP, citalopram + bupropion; CITBUS, citalopram + buspirone; CITCT, citalopram + cognitive therapy; CT, cognitive therapy; SER, sertraline; VEN, venlafaxine.
ROC curve for models with top ten predictors.
Discussion
The current study constitutes the first comprehensive effort to predict second-step treatment outcomes for MDD using ensemble machine learning algorithms. There exists relatively little prognostic information about what factors are predictive of successful pharmacotherapy and cognitive therapy for patients who do not achieve remission after an initial trial of SSRI treatment. Consistent with a precision psychiatry framework (Bernardini et al., Reference Bernardini, Attademo, Cleary, Luther, Shim, Quartesan and Compton2017), efforts such as ours can be useful for developing clinical decision-support tools to guide treatment of depression. In the current study, prediction of treatment remission was evaluated with ensemble super learner algorithms, including models trained on the full set of predictor variables and on a more parsimonious subset of the top ten predictors.
Results of the final super learner models demonstrated differential predictive performance dependent on the particular intervention. AUC values ranged from 0.51 to 0.82 for the models trained on the full predictor set. The best predictive performance was associated with the cognitive therapy (AUC = 0.82) and bupropion (AUC = 0.66) second-step treatments, whereas prediction of remission status failed to significantly outperform the chance for the remaining five treatment strategies. For cognitive therapy, both the positive predictive value (0.75) and negative predictive value (0.88) were good, indicating that a high proportion of the machine learner's predictions of remitters and non-remitters were accurate. For bupropion, the positive predictive value (0.50) indicated that roughly half of the learner's predictions of being a remitter were accurate, whereas the negative predictive value (0.75) indicted that a majority of the learner's predictions of being a non-remitter were accurate.
Of note, training super learner models on a more parsimonious subset of features, identified by a univariate filtering feature selection process, resulted in statistically significant prediction for more treatments. Again, predictive performance was best for cognitive therapy (AUC = 0.72) and bupropion (AUC = 0.70), and statistically significant prediction was also obtained for combined citalopram and cognitive therapy (AUC = 0.65), venlafaxine (AUC = 0.60), and sertraline (AUC = 0.59). For every treatment except the combined citalopram and bupropion intervention, the positive predictive values were regularly lower than the negative predictive values, indicating predictions of being a non-remitter may be more accurate than predictions of being a remitter. Submitting the models to such feature selection processes may have augmented successful prediction for more treatments, as feature selection filtering can mitigate potential overfitting and enhance model accuracy by reducing noise or redundancy in the predictor set (Kuhn & Johnson, Reference Kuhn and Johnson2013). Simulations have illustrated how including non-informative predictors can undermine model performance for various algorithms (Kuhn & Johnson, Reference Kuhn and Johnson2019). Thus, variable filtering procedures reducing non-informative predictors may have enhanced model performance in the current study.
Among all the second-step treatment options, the super learner models yielded the best performance for cognitive therapy using both the full and parsimonious predictor set. These results are broadly consistent with extant literature on moderators and predictors of psychotherapy and pharmacotherapy for depression, suggesting that traditional demographic and self-report clinical variables have attained variable success in predicting treatment outcomes in general (Kessler, Reference Kessler2018; Papakostas & Fava, Reference Papakostas and Fava2008; Perlman et al., Reference Perlman, Benrimoh, Israel, Rollins, Brown, Tunteng and Berlim2019; Spielmans & Flückiger, Reference Spielmans and Flückiger2018). With the exception of bupropion, it appears that machine learning algorithms afforded worse prediction for interventions involving pharmacotherapy. There is no obvious reason why prediction of remission status of cognitive therapy was more successful than that of pharmacotherapy. Overall, these results underscore the need to improve precision medicine approaches for pharmacotherapy treatments. Future research may consider adopting predictor sets that extend beyond self-report psychological constructs such as genetic variables, inflammatory markers, neuroimaging substrates, blood lab markers, and electronic health record data. Comprehensive reviews have identified such features as having potential prognostic value for antidepressant outcomes (Perlman et al., Reference Perlman, Benrimoh, Israel, Rollins, Brown, Tunteng and Berlim2019), and it may be profitable for machine learning models to complement traditional psychosocial variables with such predictors. Furthermore, electronic health record data has demonstrated clinical utility in predicting important clinical outcomes (e.g. first-line antidepressant response, treatment resistance, treatment dropout, etc.; Hughes et al., Reference Hughes, Pradier, Ross, McCoy, Perlis and Doshi-Velez2020; Lage, McCoy, Perlis, & Doshi-Velez, Reference Lage, McCoy, Perlis and Doshi-Velez2022; Pradier, McCoy, Hughes, Perlis, & Doshi-Velez, Reference Pradier, McCoy, Hughes, Perlis and Doshi-Velez2020; Sheu et al., Reference Sheu, Magdamo, Miller, Das, Blacker and Smoller2023).
Of the top predictors identified as part of the feature selection filters, work and social impairment, quality of life, and depression severity were the most common predictors shared across all treatment strategies. This pattern of results is consistent with treatment moderators identified in prior reviews (Papakostas & Fava, Reference Papakostas and Fava2008; Perlman et al., Reference Perlman, Benrimoh, Israel, Rollins, Brown, Tunteng and Berlim2019). Indeed, the original evaluation of first-step citalopram treatment revealed that these same constructs differentiated remission status in the STAR*D trial (Trivedi et al., Reference Trivedi, Rush, Wisniewski, Nierenberg, Warden and Ritz2006). It is noteworthy that these factors are also broadly predictive of second-step treatment outcomes across several different intervention options, attesting to their clinical utility and value.
One other striking feature of the results is that all the models were associated with high specificity and relatively modest sensitivity, irrespective of whether the models were trained on the full predictor set or filtered predictor set. Moreover, the models demonstrated better negative predictive value rates relative to positive predictive value rates. In terms of interpretability, this pattern would indicate that the predictions classified as non-remitters have greater accuracy, which fosters more confidence in instances when the algorithm predicts someone to be a non-remitter. This balance of performance metrics warrants a nuanced perspective for potential clinical implementation.
Embracing a precision medicine perspective has the potential to transform our approach to second-step care for depression. Within the overall context of machine learning in psychiatry, model performance in terms of AUC for several of the treatments in the current study was similar to that of other machine learning studies predicting treatment outcomes for depression (Chekroud et al., Reference Chekroud, Zotti, Shehzad, Gueorguieva, Johnson, Trivedi and Corlett2016; Nie, Vairavan, Narayan, Ye, & Li, Reference Nie, Vairavan, Narayan, Ye and Li2018). Notably, the current study is the first to comprehensively evaluate the utility of super learner algorithms across a wide array of second-step treatment options. Future efforts to develop accurate machine learning models predicting treatment outcomes can facilitate clinical decision-making. Machine learning algorithms can be incorporated into clinical support tools that provide clinicians with the predicted probability of remission from a given treatment, which can better inform treatment planning and discussions with patients about the relative benefits of an intervention.
A principal strength of the current study is that it examines prediction of outcomes for multiple treatments rather than prediction of a single intervention, as is commonplace for most research (Perlman et al., Reference Perlman, Benrimoh, Israel, Rollins, Brown, Tunteng and Berlim2019). By developing predicative algorithms for multiple treatments, as opposed to a single treatment, patient data can be submitted as predictive features to several different treatment algorithms simultaneously to determine the probability of remission in response to each intervention. This approach may promote a more informed decision-making process between patients and providers about what treatment protocol may be preferable, after considering a confluence of factors such as predictive modeling results for each treatment, patient preference, potential side-effects, as well as other relevant information. Obtaining the predictive probability of success for each treatment strategy better facilitates these types of more in-depth patient-provider conversations and decision-making, which has advantages over a winner-take-all approach in which machine learning is used to recommend only one treatment strategy based on the highest predictive probability value. This latter approach can obscure important information and deter thoughtful consideration of the value of these other factors that influence treatment preference.
Given the relatively higher negative predictive value performance of the models, the machine learning algorithms can afford insight into which specific second-step treatments may not confer clinical benefit. Furthermore, the results the present study are noteworthy in light of the fact that models were trained only on self-reported demographic and clinical symptom data, which represents a more feasible approach in comparison to using more expensive and burdensome data sources for machine learning models such as neuroimaging or biomarker data (Lee et al., Reference Lee, Ragguett, Mansur, Boutilier, Rosenblat, Trevizol and McIntyre2018).
Notwithstanding the strengths of the current study, certain limitations warrant mention. First, although some models evinced good predicative performance, AUC values were not statistically better than chance in predicting remission status for several treatments. Thus, it would be contraindicated to recommend the current models for clinical use in their current form, as the predictions from certain models suffer from poor accuracy and may incur a potential risk of misclassification. Given that the current study leveraged the largest clinical trial of second-step depression treatments, the current results provide an important evidence base about the strengths and limitations of machine learning prediction models in this setting. Future research may be able to improve model accuracy by considering other potential predictors that were not available in the STAR*D study (e.g. genetic factors, neuroimaging substrates, electronic health record data, etc.). Second, the sample size for some treatment conditions was relatively modest in the context of machine learning research. That notwithstanding, the STAR*D trial represents the most extensive and well-powered study examining second-step treatments for depression, and machine learning studies have resulted in success using smaller sample sizes (e.g. Flygare et al., Reference Flygare, Enander, Andersson, Ljótsson, Ivanov, Mataix-Cols and Rück2020). To mitigate this, we applied nested cross-validation, which can reduce bias in models with limited sample sizes. Third, for three of the treatment conditions, the remission status outcome was imbalanced and adversely influenced model performance. To address this, we used the SMOTE procedure to produce a more balanced class profile for model training. Fourth, results might be influenced by the nature of randomization for the STAR*D trial, as patients willing to accept randomization to different treatment strategies may exhibit differences that could influence the difference across the treatment groups.
The current study constitutes the most comprehensive examination to date of super learning ensemble models in predicting remission across an extensive variety of different second-step treatments for depression. Before translating predictive modelling to clinical practice, improvements in model performance will be needed and might be achieved through using a more diverse array of multimodal predictors (e.g. self-report, biomarker data, etc.). In conclusion, the current study illustrates a proof-of-concept approach using machine learning as a tool for improving personalized second-step treatment for depression and provides insight into important next steps for future precision medicine research on depression.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291724000497
Acknowledgements
The authors would like to commend Dr Myles Lewis for his valuable assistance with the R package ‘nestedcv’, and to thank Dr Christopher Kennedy for his conversations about machine learning.
Funding statement
The study received financial support from the Kaplen Fellowship on Depression.
Competing interests
The authors declare no competing interests.
