Background: Outcome and adverse events are the two primary factors when planning a safe and successful antipsychotic therapy. However, evidence is scarce regarding the prediction of outcome and adverse events in an individual patient.
Objective: Our aim is to predict the outcome and adverse events in antipsychotic treatment.
Study Design: We plan to predict the outcome and adverse events by implementing a machine learning approach in an individual participant data meta-analysis of randomized controlled trials (RCTs).
Participants: Schizophrenia, bipolar disorder, and schizoaffective disorder.
Main outcome measure: Our main outcome measure will be reduction of major symptoms (i.e. psychosis or mania).
Statistical analysis: Non-linear regression of the above-mentioned outcome measures will be carried out using state-of-the-art artificial neural networks. The model’s accuracy will be compared with alternative approaches including linear regression models and support vector regression.
Providing the appropriate antipsychotic substance in psychiatric disorders is a complex process that involves prediction of at least two key factors: outcome and adverse events1. The desired outcome should ideally outweigh potential adverse events. However, in clinical routine, prediction of these two factors in individual remains elusive2.
Predictors of outcome have been investigated in different clinical, social, and genetic domains2. Focussing on the clinical history, Kinon et al. investigated early response to an antipsychotic medication as a predictive factor for a later response3. Using data from five randomized controlled trials (n = 1077 schizophrenia patients), they showed that early non-response was a robust predictor of continued later lack of response3. Similarly, social factors likely influence the functional outcome in antipsychotic treatment. For instance, Köhler-Forsberg et al. reported that living with a partner was the strongest predictor of social functioning (assessed by Global Assessment of Functioning – GAF) after clozapine initiation in schizophrenia patients4. Pharmacogenetic investigations showed that a combination of six polymorphisms in neurotransmitter-receptor-related genes resulted in a significant 76.7% prediction of clozapine response and a sensitivity of 95% for satisfactory response5. Taken together, these findings indicate that multifactorial variables from different domains contribute to the outcome of antipsychotic treatment.
In addition to outcome, prediction of adverse events (AEs) is the other key factor when planning a safe and successful therapy. Even though AEs are frequent in antipsychotics6, little is known about their prediction according to individual patient characteristics7. Polypharmacy is a known risk factor for the occurrence of adverse events8. Furthermore, the risk for increased weight gain during treatment with olanzapine was reported to be threefold in subjects with at least one allele at each locus of leptin and leptin receptor8. Polymorphisms (A1 allele) of the dopamine D2 receptor gene Taq1 in females were associated with increased prolactine levels during treatment with bromperidol9.
The scarce literature highlights the need for research to optimize prediction of outcome and adverse events in antipsychotic treatment. Importantly, both measures depend on multiple variables (e.g. demographic, genetic, and clinical) most likely in a complex non-linear interaction, which cannot be captured in conventional linear regression models, where the dependent variable is predicted as a weighted sum of individual predictors. Neural networks represent the most advanced technique to tackle such non-linear regression problems. Employing these techniques, we aim at identifying complex patterns of predictors of treatment outcome and adverse events. This study could have a major impact on health of patients and help to identify crucial predictors of outcome and adverse events and may help to promote the development of personalized and precise therapeutic strategies.
Primary objective: A. Identify patterns of predictors for outcome during antipsychotic treatment.
Secondary objective: B. Identify patterns of predictors for adverse events during antipsychotic treatment.
Endpoints: The following two groups of endpoints (1A-4A and 5B-10B) are clustered according to the two groups of objectives (A and B).
1A. Pattern of predictive variables for symptom reduction during antipsychotic treatment (all antipsychotics pooled).
2A. Pattern of predictive variables for symptom reduction during antipsychotic treatment (assessed for each antipsychotic individually).
3A. Pattern of predictive variables for increase in global functioning during antipsychotic treatment (all antipsychotics pooled).
4A. Pattern of predictive variables for increase in global functioning during antipsychotic treatment (assessed for each antipsychotic individually).
5B/6B/7B. Pattern of predictive variables for occurrence/severity/duration of adverse events during antipsychotic treatment (all antipsychotics pooled).
8B/9B/10B. Pattern of predictive variables for occurrence/severity/duration of adverse events (assessed for each antipsychotic individually).
Participant-level data provided from randomized controlled trials (RCTs) on antipsychotic treatment in patients with schizophrenia, bipolar disorder, and schizoaffective disorder will be included. All routes of antipsychotic administration (e.g. oral and injection) will be included. The primary endpoint is six weeks treatment duration but durations from 4 to 12 weeks will also be included10.
The main outcome is reduction in the total score of major symptoms (psychosis or mania) from baseline to endpoint of six (four to twelve) weeks post-baseline. All assessment time points in this timeframe will be included. We aim to implement the same score for each disorder:
• Psychosis: Positive and Negative Syndrome Scale (PANSS)11
• Mania: Young Mania Rating Scale (YMRS)12
Oral and long-acting injectable antipsychotics will be calculated separately.
Potential predictors include variables derived from demographic data, clinical examinations, and laboratory investigations. Specifically, we aim at determining predictive combinations (patterns) of these predictors using artificial neural networks. We will investigate if these predictive combinations are similar for all investigated psychiatric disorders / medications or unique to specific disorders / medications.
We will include additional variables / characteristics potentially associated with outcome and AEs. Global functioning (endpoint 3A and 4A) will be measured on the Clinical Global Impression (CGI) – scale13. Occurrence of AEs (Yes / No) will be measured according to trial documentation. Duration of AEs is defined as cumulative number of days the AE occurred. Severity of AEs is measured as mild / moderate / severe.
We will merge individual patient data from the RCTs provided by The YODA Project. Separate analyses will be made for each diagnosis (i.e. schizophrenia, bipolar disorder, and schizoaffective disorder). We will use Deep Learning neural network methods with emphasis on uncertainty quantification and robustness against outliers. Uncertainty quantification does not just allow to predict the expected treatment outcome and risk for adverse events, but also confidence intervals of the prediction14,15. To choose a specific treatment option together with a patient in the sense of informed consent and shared decision making, it is crucial to have a measure of the prediction's certainty on hand. Hence, applying and refining techniques for certainty quantification of individualized treatment predictions might constitute a substantial advance on the road to individualized treatment recommendations based on multiple patient characteristics from different domains. Outlier robustness is important since real-life data is always noisy and contains contaminated learning data, which might impede the proper learning of predictive patterns.
We plan to compare the predictions of the neural networks with other established machine learning regression techniques such as support vector regression and random forests. We will construct multiple test-train folds for repeated cross-validation through partition of the original cohort into a subset for training purposes and a subset for testing16. Model performance is captured by the root mean squared error (e.g. deviance between predicted and real treatment outcome) and data likelihood and will be examined in an independent cohort subset.
Missing data will be treated as recommended by Little et al.17: First we will register if reasons for missing data were documented and develop a primary set of assumptions about the cause for missing data17. Then the primary set of assumptions will be followed by multiple imputation by chained equations and robustness tested with a sensitivity analysis17.
The two main factors guiding the choice of a specific antipsychotic treatment are its outcome and risk of adverse events. A personalized therapeutic strategy would be based on prediction of these factors, where the treatment effect should ideally outweigh potential adverse events. However, in clinical routine, prediction of these two factors remains elusive. We intend to predict outcome and adverse events in patients with schizophrenia, schizoaffective disorder, and bipolar disorder treated with antipsychotics in randomized controlled trials using a machine learning approach trained on individual demographic, clinical, and laboratory parameters.
• Milestone 1 at 0 months: Data preparation and implementation of Deep Learning neural network methods starts.
• Milestone 2 at 12 months: Analysis of objective A starts.
• Milestone 3 at 24 months: Analysis of objective B starts.
• Milestone 4 at 36 months: Analyses of objectives are completed and papers drafted.
The YODA project will be informed about the completion of each milestone and reports will be made available.
To benefit both health professionals and patients we will present the study at internationally accredited conferences (e.g. symposia at the WPA) and make the study available in major medical journals (e.g. JAMA Psychiatry, American Journal of Psychiatry, Lancet Psychiatry).
1. Leucht, S. et al. Comparative efficacy and tolerability of 15 antipsychotic drugs in schizophrenia: a multiple-treatments meta-analysis. Lancet 382, 951–962 (2013).
2. Stern, S., Linker, S., Vadodaria, K. C., Marchetto, M. C. & Gage, F. H. Prediction of response to drug therapy in psychiatric disorders. Open Biol. 8, 180031–13 (2018).
3. Kinon, B. et al. Predicting response to atypical antipsychotics based on early response in the treatment of schizophrenia☆. Schizophrenia Research 102, 230–240 (2008).
4. Köhler-Forsberg, O., Horsdal, H. T., Legge, S. E., MacCabe, J. H. & Gasse, C. Predictors of Nonhospitalization and Functional Response in Clozapine Treatment. Journal of Clinical Psychopharmacology 37, 148–154 (2017).
5. Arranz, M. J. et al. Pharmacogenetic prediction of clozapine response. Lancet 355, 1615–1616 (2000).
6. Citrome, L. A Review of the Pharmacology, Efficacy and Tolerability of Recently Approved and Upcoming Oral Antipsychotics: An Evidence-Based Medicine Approach. CNS Drugs 27, 879–911 (2013).
7. Plesnicar, B. K. Personalized antipsychotic treatment: the adverse effects perspectives. Psychiatr Danub 22, 329–334 (2010).
8. Kadra, G. et al. Predicting parkinsonism side-effects of antipsychotic polypharmacy prescribed in secondary mental healthcare. Journal of Psychopharmacology 32, 1191–1196 (2018).
9. Mihara, K. et al. Relationship between Taq1 A dopamine D2 receptor (DRD2) polymorphism and prolactin response to bromperidol. Am. J. Med. Genet. 105, 271–274 (2001).
10. Samara, M. et al. Initial symptom severity of bipolar I disorder and the efficacy of olanzapine: a meta-analysis of individual participant data from five placebo-controlled studies. The Lancet Psychiatry 4, 859–867 (2017).
11. Kay, S., Fiszbein, A. & Opler, L. A. The Positive and Negative Syndrome Scale (PANSS) for schizophrenia. Schizophrenia Bulletin 2, 261–276 (1987).
12. Young, R. C., Biggs, J. T., Ziegler, V. E. & Meyer, D. A. A rating scale for mania: reliability, validity and sensitivity. The British Journal of Psychiatry 133, 429–435 (1978).
13. Guy, W. ECDEU assessment manual for psychopharmacology. (Rockville, Md., USA : U.S. Dept. of Health, Education, and Welfare, Public Health Service, Alcohol, Drug Abuse, and Mental Health Administration, National Institute of Mental Health, Psychopharmacology Research Branch, Division of Extramural Research Programs, 1976).
14. Gurevich, P. & Stuke, H. Gradient conjugate priors and multi-layer neural networks. arXiv:1707.07287v3 [stat.ML] 1–44 (2018).
15. Gurevich, P. & Stuke, H. Pairing an arbitrary regressor with an artificial neural network estimating aleatoric uncertainty. arXiv:1707.07287v3 [stat.ML] 1–29 (2018).
16. Chekroud, A. M. et al. Cross-trial prediction of treatment outcome in depression: a machine learning approach. The Lancet Psychiatry 3, 243–250 (2016).
17. Little, R. J. et al. The Prevention and Treatment of Missing Data in Clinical Trials. N Engl J Med 367, 1355–1360 (2012).