Introduction
Major depressive disorder is a serious, disabling, life-shortening illness with high lifetime risks. Present practice entails a trial-and-error approach with beginning a medication, adjusting doses upwards (side effects permitting), and then deciding (after 2–8 weeks) to either stop or alter the treatment if response has not occurred (APA, 2010).
If we could reliably predict that non-response will be present at 8 weeks, within a few weeks into treatment, we could reduce exposure of selected patients' treatment that is very likely to not succeed. Such a prediction could save patients and care systems the time, side-effect risk and costs associated with a more prolonged treatment trial that in turn could hasten treatment revisions for such patients.
While several efforts to forecast response or non-response at 8–12 weeks using pretreatment features alone or with post-baseline symptom-change measures, none has been of sufficient value to become part of routine care (Nierenberg et al. Reference Nierenberg, McLean, Alpert, Worthington, Rosenbaum and Fava1995; Quitkin et al. Reference Quitkin, Petkova, McGrath, Taylor, Beasley, Stewart, Amsterdam, Fava, Rosenbaum, Reimherr, Fawcett, Chen and Klein2003). The statistical procedures used to date have not directly led to a clinical decision because they have not aimed to control the positive predictive value (PPV) (the probability that a patient who is predicted to respond actually does so) or the negative predictive value (NPV) (the probability that a patient who is predicted to not respond actual does so). NPV and PPV are more relevant to clinical decision making than sensitivity and specificity because clinicians have to try to decide for each individual whether to continue or stop (alter) the treatment. They must have a reasonably high degree of certainty about the decision before they act.
However, unlike traditional statistical approaches such as logistic regression and classification trees (Venables & Ripley, Reference Venables and Ripley1994; Trevor et al. Reference Trevor, Tibshirani and Friedman2006), recursive subsetting allows binary predictions (will or will not respond later) in which the clinician can set the degree of certainty needed for treatment. Specifically, both logistic regression and classification trees cannot allow investigators to control the NPV and PPV directly. These approaches are thus not helpful for the current goal. Typically, researchers will just report PPVs and NPVs as they are rather than trying to control them at some prescribed values. For example, Henkel et al. (Reference Henkel, Seemüller, Obermeier, Adli, Bauer, Mundt, Brieger, Laux, Bender, Heuser, Zeiler, Gaebel, Mayr, Möller and Riedel2009) reported an NPV of 0.37 for one of their classification rules, which is clearly not good enough. Since the full benefit of treatment takes 2–3 months to realize (Trivedi et al. Reference Trivedi, Rush, Wisniewski, Nierenberg, Warden, Ritz, Norquist, Howland, Lebowitz, McGrath, Shores-Wilson, Biggs, Balasubramani and Fava2006), we could use recursive subsetting with a predefined reasonably high NPV (e.g. 0.75, 0.80, 0.90) so that the clinician can act.
We have previously found that recursive subsetting allows clinicians to select the degree of certainty that he/she needs to identify individuals who will do poorly (e.g. not respond) at 6 weeks using baseline features and one post-baseline symptom-change measure (Kuk et al. Reference Kuk, Li and Rush2010). In practice, however, multiple post-baseline treatment visits are held at which additional post-baseline symptom-change measures can be obtained. This report uses a recursive subsetting approach with a new dataset in which we evaluate the utility of baseline feature measurements and one or two follow-up time points (at weeks 2 and 4) to predict non-response at 8 weeks. If successful, then efforts to automate this approach for clinical use would be justified.
Specifically, this report addresses three questions:
(1) Can recursive subsetting produce clinically useful and actionable information in a new dataset of representative out-patients with major depressive disorder?
(2) Does the best actionable information depend on both baseline and post-baseline findings?
(3) What are the advantages of using multiple time points over a single time point in predicting patient response? How do we make use of the time course/trajectory of early symptom change to predict eventual outcome?
Method
Study overview
This report used data from the Combining Medications to Enhance Depression Outcomes (CO-MED) trial (www.co-med.org;; Rush et al. Reference Rush, Trivedi, Stewart, Nierenberg, Fava, Kurian, Warden, Morris, Luther, Husain, Cook, Shelton, Lesser, Kornstein and Wisniewski2011), a 7-month single-blinded, randomized, placebo-controlled trial that compared the efficacy of each of two different antidepressant medication combinations (bupropion-sustained release+escitalopram and venlafaxine-extended release+mirtazapine) versus escitalopram+placebo in 12 weeks of first-step medication treatment for major depressive disorder. Study details and results are available elsewhere (Rush et al. Reference Rush, Trivedi, Stewart, Nierenberg, Fava, Kurian, Warden, Morris, Luther, Husain, Cook, Shelton, Lesser, Kornstein and Wisniewski2011). In brief, clinical research coordinators collected standard baseline sociodemographic and clinical information, as well as the 16-item Quick Inventory of Depressive Symptomatology – Self-Report (QIDS-SR16) (Kornstein et al. Reference Kornstein, Sloan and Thase2002; Rush et al. Reference Rush, Trivedi, Ibrahim, Carmody, Arnow, Klein, Markowitz, Ninan, Kornstein, Manber, Thase, Kocsis and Keller2003, Reference Rush, Bernstein, Trivedi, Carmody, Wisniewski, Mundt, Shores-Wilson, Biggs, Woo, Nierenberg and Fava2006; Trivedi et al. Reference Trivedi, Rush, Ibrahim, Carmody, Biggs, Suppes, Crismon, Shores-Wilson, Toprac, Dennehy, Witte and Kashner2004) to measure depressive symptom severity. During the initial 12 weeks of acute treatment, clinic visits were recommended at baseline (week 0) and at weeks 1, 2, 4, 6, 8, 10 and 12. For this report, we defined response as a ⩾50% reduction from baseline in the total QIDS-SR16 score by week 8. (Note: we chose the week 8 time point because CO-MED allowed participants with minimal response at that time to exit.)
Sample selection
For inclusion in these analyses, participants had to have complete QIDS-SR16 data at baseline and at week 2, 4 and 8. The analysis included 447 participants. After participants were provided with a complete description of the study, written informed consent was obtained.
Statistical analysis
For these analyses, we used the recursive subsetting approach (Kuk et al. Reference Kuk, Li and Rush2010) and we chose an initial level of 75% NPV as a requisite for a clinically useful, reliable prediction.
We selected three dichotomous baseline features based on previous reports (Kornstein et al. Reference Kornstein, Sloan and Thase2002; Trivedi et al. Reference Trivedi, Rush, Wisniewski, Nierenberg, Warden, Ritz, Norquist, Howland, Lebowitz, McGrath, Shores-Wilson, Biggs, Balasubramani and Fava2006; Fava et al. Reference Fava, Rush, Alpert, Balasubramani, Wisniewski, Carmin, Biggs, Zisook, Leuchter, Howland, Warden and Trivedi2008) that might predict treatment non-response: gender, anxious features (Hamilton anxiety/somatization factor score ⩾7) (Rush et al. Reference Rush, Fava, Wisniewski, Lavori, Trivedi, Sackeim, Thase, Nierenberg, Quitkin and Kashner2004) and chronic index episode (episode duration >2 years) and two continuous variables (baseline severity based on QIDS-SR16 and age). (Note that a variety of other baseline variables could also have been used.)
We used change in QIDS-SR16 from baseline to week 2 and baseline to week 4 to predict non-response at week 8. Specifically, we defined the proportion of reduction in QIDS-SR16 score from baseline to week 2 (W 2) using the formula W 2=(S 0−S 2)/S 0, with S 0 being the participant's baseline QIDS-SR16 score and S 2 being the participant's QIDS-SR16 score at week 2. Similarly, we defined the proportion of reduction in QIDS-SR16 score from baseline to week 4 (W 4) using the formula W 4=(S 0−S 4)/S 0, with S 4 being the participant's QIDS-SR16 score at week 4.
For W 2, we specified six non-overlapping ascending intervals: W 2<0, 
, 
, 
, 
and 
. These cut-offs were selected on the basis of linear extrapolation of response and resulted in six meaningful categories of participants, namely those who were worse off after 2 weeks (W 2<0), those with very little improvement after 2 weeks (
), those who had modest symptom improvement but were still off pace to respond by week 8 (), those right on course to respond by week 8 (), those ahead of course to respond by week 8 (), and, those who had already responded by week 2 (). We expected that lower values of W 2 would be more closely associated with a non-response. Therefore, there might be a decreasing trend in the probability of not responding to treatment from the 1st to the 6th group. We started with the group least likely to respond (category 1) and partitioned the participants into subsets according to baseline information such that a satisfactory NPV could be achieved while the sample size of the subset is maximized. The same procedure was applied to the following groups sequentially. We stopped the partitioning when the overall NPV for the subsets reached 75% and further subsetting did not increase the NPV (Kuk et al. Reference Kuk, Li and Rush2010).
We also used appropriately modified cut-offs for W 4 to define six intervals of W 4: W 4<0, 
, 
, 
, 
, and 
with the same interpretation as before. We could then conduct the same recursive subsetting program described above for W 2. Since week 4 is temporally closer to week 8 than to week 2, we expected that week 4 changes might be more predictive of the final outcome. We used W 4 information in three different ways: (i) using W 4 alone; (ii) using W 2 alone in the first subsetting procedure and then using W 4 alone in a second subsetting on those for whom we could not predict with the first subsetting; (iii) using W 4 and W 2 together in a single subsetting procedure.
We further used the recursive subsetting method on two different subgroups (severe/non-severe groups and chronic/non-chronic). For these subgroup analyses, we used an NPV of 80% to indicate a clinically useful reliable prediction.
One important statistical issue is to assess the external validity of our procedure. We consider the cross-validation approach in this paper. By splitting the whole data into 10 subsets at random, we formed separate training samples and test samples and evaluated the accuracy of the test samples by using the prediction rule constructed from the training samples. The cross-validated NPV provides a reasonable estimate for the prediction accuracy for future analysis.
Results
Predictions of non-response with baseline data alone
Of the 447 participants, 250 (56%) responded by week 8. Table 1 shows the proportion of participants who did and did not respond by week 8 for every possible combination of the three dichotomous baseline variables: chronicity, gender and anxious features. None of these baseline measures were clinically useful in predicting either response or non-response at 8 weeks because none of the PPVs and NPVs reached the pre-specified required level of accuracy of 75%.
Table 1. Proportions of participants not responding and responding to antidepressant medication at week 8 for every combination of three dichotomous baseline variables
a Denominator=number of participants with the designated combination of baseline features. Numerator=number of participants with non-response at week 8 from the group.
b Denominator=number of participants with the designated combination of baseline features. Numerator=number of participants with response at week 8 from the group.
Fig. 1 shows the empirical NPV curves for age and baseline QIDS-SR16 scores, each a continuous function of the percentiles of the continuous predictor. Neither age nor severity was sufficiently predictive for clinical purposes since the NPV value is <0.7 for almost the entire range of the cut-off. Taken together, Table 1 and Fig. 1 indicate that baseline information alone is not sufficient to predict non-response by week 8.
Fig. 1. Negative predictive value (NPV) curves for age, baseline 16-item Quick Inventory of Depressive Symptomatology – self-rated (QIDS-SR16) score, and percentage reduction in QIDS-SR16 score at week 2.
Prediction of non-response using week 2 data
Fig. 1 also shows that the NPV curve for predicting non-response by week 8 based on percentage reduction in QIDS-SR16 score from baseline to week 2 (W 2) is above the age and baseline QIDS-SR16 curves for almost the entire percentage range. Thus, percentage reduction from baseline to week 2 (W 2) in QIDS-SR16 score is far more informative about week 8 outcome than either age or baseline severity.
To further evaluate the predictive usefulness of W 2, we formed the six categories noted above: W 2<0, 
, , , , and . The proportion of participants who did not respond by 8 weeks in each category was 0.64, 0.68, 0.64, 0.60, 0.38 and 0.22, respectively. While NPV was relatively higher when W 2 was small, each NPV value was still <0.75. Consequently, the clinician cannot decide to stop the medication because the certainty is insufficient.
We then incorporated baseline features into the W 2 categories and conducted recursive subsetting to identify subsets of participants with a high NPV within each W 2 category. One subset did achieve the NPV threshold of 0.75. Specifically, participants whose QIDS-SR16 had not dropped by at least 1/8 (
) by week 2 and who were in a chronic index episode were predicted to not respond by week 8. In fact, 48 participants were in this group, and 36 of them were true non-responders at week 8, which yielded an overall NPV of 36/48=0.75.
Prediction of non-response using week 4 data
Fig. 2 displays the NPV curve for predicting who will not respond to treatment by week 8 based on the percentage reduction in QIDS-SR16 score from baseline to week 2 and from baseline to week 4. The NPV curve for W 4 is above the W 2 curve for most of its range, which indicates better predictive power for W 4.
Fig. 2. Negative predictive value (NPV) curves for percentage reduction in 16-item Quick Inventory of Depressive Symptomatology – self-rated (QIDS-SR16) score at week 2 and at week 4.
We then considered the six categories of W 4 described earlier: W 4<0, , , , , and . The proportion of participants who did not respond in each category was 0.85, 0.67, 0.57, 0.63, 0.35 and 0.22, respectively. The first category achieved the required NPV.
We then conducted recursive subsetting to seek further subsets of the remaining five categories and arrived at two subsets: (i) W 4<0 (i.e. worsening by week 4); (ii) 
and being in a chronic index episode. We could then make a ‘will not respond’ prediction for 79 participants, 60 of whom were true non-responders giving an NPV of 76% (60/79). Using baseline and W 4 data allowed us to make actionable predictions for more participants than relying only on baseline and W 2 data. Of these 79 participants, 50 were not predicted by using the baseline and W 2 information. On the other hand, there were 19 participants predicted by W 2 but not by W 4. The fact that W 2 and W 4 predicted quite different individuals suggests that it is better to use both W 2 and W 4 to make predictions than to use just one of them.
Prediction of non-response using baseline, week 4 and week 2 data
To further evaluate this idea, we incorporated a two-step sequential approach. For step 1, we predicted 48 participants would not respond by using W 2 and baseline data as described in the preceding subsection. For step 2, we used W 4 and baseline data for the remaining 399 (447–48) participants. For this group, we predicted that another 39 would not respond. They belonged to the following subsets: (i) W 4<0; (ii) and anxious features. Thus, overall we identified 87 participants who we predicted would not respond by week 8, of whom 66 were true non-responders (NPV of 75%). Thus, this combination could predict poor outcomes for nearly 20% of the sample (87/447). Using both W 2 and W 4 data in a two-step sequence, along with baseline information, provided predictions on more participants than using either one alone.
If a higher NPV of 80% is desired, the first stage of the sequential procedure based on W 2 and baseline features is not able to predict any non-responder with the required certainty of 80% or more. Thus no action can be taken at week 2. At the second stage at week 4, both W 4 and W 2 become available, and we can use both of them as well as baseline features to predict who are the non-responders. As shown in Fig. 3, we are able to predict 67 non-responders (approximately 16% of the sample) from the following categories: (i) W 4<0 (26 participants, of whom 22 are true non-responders); (ii) and chronic episode (25 participants, of whom 19 are true non-responders); (iii) and (16 participants, of whom 14 are true non-responders). Using W 4 alone, recursive subsetting identified only the first two categories. Thus making use of both W 2 and W 4 allowed us to uncover the third subset of 16 participants as likely non-responders. There is an interesting way to describe this group of participants. They are participants who had not kept up their progress. While ‘‘ is considered ‘on course’ in week 2, if there is no further reduction in QIDS-SR16 score and the participant is ‘‘ after 4 weeks, then this stagnation of progress is a sign that a patient may not respond to treatment. This is an example of how using both W 2 and W 4 gives us extra mileage in predicting patient response than using either alone. In this case, the cross-validated NPV from a 10-fold cross-validation procedure is 0.809. Results of cross-validation in all other cases show that the test samples maintain the desired NPV and suggest that our results are applicable for future data.
Fig. 3. Recursive subsetting algorithm using week 2 and week 4 data with a target negative predictive value (NPV) of 0.8. The numbering of subsets is according to our previous report (Kuk et al. Reference Kuk, Li and Rush2010), indicating the sequence of the subsets generation. The NPV of each selected subset is shown in bold and CumNPV is the cumulative NPV for all the selected subsets.
Application of recursive subsetting to clinically defined subgroups
We evaluated two different subgroups, defined by baseline features, to determine if the approach might be particularly useful in one or another subgroup.
We divided the participants into severe (baseline QIDS-SR16 >15) and non-severe (baseline QIDS-SR16 ⩽15) groups, fixed the NPV level at 80%, and conducted recursive subsetting for these two groups using W 2, W 4 and baseline measures together. For the severe group (n=216), we could predict 35 non-responders (16% of the sample) who were from the following subsets: (i) W 4<0 (n=8); (ii) and chronic episode (n=11); (iii) and chronic episode (n=16). There were 30 true non-responders, giving an NPV of 85%. For the non-severe group (n=231), we could predict 26 non-responders (11% of the sample) who were from the following subsets: (i) W 4<0 (n=18); (ii) and (n=8). There were 22 true non-responders, giving an NPV of 84%. The results from the two subgroups were not that different.
We also conducted separate analyses for the chronic/non-chronic groups, with an NPV level at 80% using recursive subsetting. For the chronic group (n=244), we could predict 50 non-responders (20% of the sample) who were from the following categories: (i) W 4<0 (n=15); (ii) and anxious features (n=18); (iii) and 
(n=17). There were 42 true non-responders, giving an NPV of 84%. For the non-chronic group (n=203), we could only predict 18 participants (only 7.2% of the sample) who were from the following categories: (i) W 4<0 (n=11); (ii) and (n=7). There were six true non-responders, giving an NPV of 86%. In this case, we could make reliable predictions for more chronic than non-chronic participants.
Discussion
The present results replicate and extend our initial work (Kuk et al. Reference Kuk, Li and Rush2010) with recursive subsetting. We found that one can make clinically useful and actionable early predictions about individual patients who would not respond later (in this case at 8 weeks) to antidepressant medication treatment. This report used 75% as the requisite NPV for the whole sample and 80% for the subgroup analyses because we thought that such degrees of certainty would be clinically meaningful. We found that this approach did identify a clinically meaningful proportion of participants early in treatment with sufficient certainty that clinicians could take action (i.e. with a 75–80% chance of non-response, treatment could be changed). This report also found that the combination of both baseline and post-baseline (i.e. depressive symptom change) information provided reliable predictions of who would not respond, while baseline information alone was not clinically useful (i.e. NPV less than 75%) so medication would not be stopped. Further, two post-baseline assessments of symptom change (weeks 2 and 4) enhanced the number of participants for whom actionable predictions could be made as compared with only baseline plus week 2 or baseline plus week 4 information. In particular, we demonstrated that stagnation of progress after some initial improvement is a potentially useful indicator of subsequent non-response to treatment.
This report focused on NPV because we wished to predict non-response. This procedure can be used to predict positive outcomes (e.g. response) via the PPV.
In addition, the recursive subsetting approach enables clinicians to specify the requisite accuracy level (e.g. NPV ⩾75% or 0.75). The choice of the accuracy level may be affected by practical considerations such as side effect risk, history of treatment resistance, cost, etc. For example, a higher NPV may be preferred if there are few treatment options, so only the patients who have less than a 10% or even 5% chance of responding would be identified to have their treatment changed.
The application of recursive subsetting is new. There are no reports other than our own (Kuk et al. Reference Kuk, Li and Rush2010) by which to compare directly our findings. As in our initial report (Kuk et al. Reference Kuk, Li and Rush2010), we used cross-validation to assess the out-of-sample performance of our procedure and observed that the stated NPV holds up well. In retrospect, this is not surprising since the recursive subsetting procedure only maximizes the number of predictions made subject to the chosen NPV rather than maximizing the NPV itself. A rigorous proof that the selected subset achieves the stated NPV with probability approaching 1 as sample size increases has been found and will be included in a more mathematical paper, together with other theoretical results. Study limitations include the nature of the sample (selected to have recurrent or chronic depression), the choice of categories to describe post-baseline changes, and the limited number of baseline measures evaluated. As well, response was defined by self-report and was based on symptom change. We would suggest, however, that other approaches such as sensitivity and specificity are not appropriate for clinical decision making since they can only evaluate overall misclassification performance. That is, they do not provide sufficiently specific predictions for individual participants so that clinicians can act.
It should be noted that the choices of cut-off points for W 2 and W 4 in our subsetting procedure were based on an assumption that symptom change follows a linear pattern. One can, however, use other (non-linear) patterns, which would lead to other cut-off values that may be more (or less) predictive. A grid search of optimal cut-off values may be employed to refine our algorithm.
Further, it should be noted that recursive subsetting can be adapted to any treatment or disease in which (a) the ultimate results/outcomes are removed in time from baseline (treatment initiation), (b) there is heterogeneity of response (i.e. some people respond and others do not), and (c) there may be a post-baseline indicator of effect – whether symptoms, laboratory test values or other indicators. In addition, the approach can be used for situations in which one is concerned about the development of longer-term side effects, and in which individual predictions are also needed (e.g. weight gain over time with selected antidepressants).
We stress that decision tree analytic techniques have also appeared in similar clinical contexts (see Mulsant et al. Reference Mulsant, Houck, Gildengers, Andreescu, Dew, Pollock, Miller, Stack, Mazumdar and Reynolds2006; Andreescu et al. Reference Andreescu, Mulsant, Houck, Whyte, Mazumdar, Dombrovski, Pollock and Reynolds2008). These findings converge with ours, as both document the importance of early symptom change during treatment of depression as a clinically actionable signal about the likelihood of non-response by 12 weeks. Taken together, our adult sample and their older adult sample (Mulsant et al. Reference Mulsant, Houck, Gildengers, Andreescu, Dew, Pollock, Miller, Stack, Mazumdar and Reynolds2006; Andreescu et al. Reference Andreescu, Mulsant, Houck, Whyte, Mazumdar, Dombrovski, Pollock and Reynolds2008) provide ample empirical evidence of the clinical utility of statistical techniques such as recursive subsetting and decision tree analysis as dynamic tools for the clinicians.
To bring this methodology into practice, an automated algorithm is needed in which a range of baseline and post-baseline variables can be provided and the requisite NPV and PPV levels can be chosen. In healthcare systems in which such measures are obtained, the deployment of such a tool would then need an evaluation as to clinical utility and cost-effectiveness.
In summary, recursive subsetting is an approach that can define and identify early individual depressed patients who are sufficiently unlikely to respond, so that clinicians can modify the treatment early in a meaningful proportion of patients. A combination of baseline (sociodemographic or clinical features) and early post-baseline depressive symptom changes provide a sufficiently accurate prediction that from 1 in 5 to 1 in 7 depressed out-patients can be spared at least 4 weeks of an ultimately ineffective treatment, which is a clinically and fiscally meaningful result.
Acknowledgements
This project was funded by the National Institute of Mental Health (NIMH) under contract N01MH90003 to UT Southwestern Medical Center at Dallas, TX, USA (principal investigator A.J.R.). The content of this publication does not necessarily reflect the views or policies of the Department of Health and Human Services, nor does mention of trade names, commercial products or organizations imply endorsement by the US Government. The NIMH had no role in the drafting or review of the manuscript, nor in the collection or analysis of the data.
We appreciate the support of Forest Pharmaceuticals Inc., GlaxoSmithKline, Organon Inc., and Wyeth Pharmaceuticals in providing medications at no cost for this trial.
The authors thank the CO-MED research team and the Depression Trials Networks for providing the data for this report. We also acknowledge the editorial support of Jon Kilner, MS, MA (Pittsburgh, PA, USA).
[The study's ClinicalTrials.gov registration no. is NCT00590863 (http://www.clinicaltrials.gov/ct2/show/NCT00590863?term=CO-MED&rank=1).]
Declaration of Interest
A.J.R. has received consulting fees from Otsuka, University of Michigan and Brain Resource; consultant/speaker fees from Otsuka and Bristol–Myers Squibb and author royalties from Guilford Publications, Healthcare Technology Systems and the University of Texas Southwestern Medical Center and has received research support from the National Institute of Mental Health.
