Leveraging stacked classifiers for exploring the role of hedonic processing between major depressive disorder and schizophrenia

Abstract

Background: Anhedonia, a transdiagnostic feature common to both Major Depressive Disorder (MDD) and Schizophrenia (SCZ), is characterized by abnormalities in hedonic experience. Previous studies have used machine learning (ML) algorithms without focusing on disorder-specific characteristics to independently classify SCZ and MDD. This study aimed to classify MDD and SCZ using ML models that integrate components of hedonic processing.

Methods: We recruited 99 patients with MDD, 100 patients with SCZ, and 113 healthy controls (HC) from four sites. The patient groups were allocated to distinct training and testing datasets. All participants completed a modified Monetary Incentive Delay (MID) task, which yielded features categorized into five hedonic components, two reward consequences, and three reward magnitudes. We employed a stacking ensemble model with SHapley Additive exPlanations (SHAP) values to identify key features distinguishing MDD, SCZ, and HC across binary and multi-class classifications.

Results: The stacking model demonstrated high classification accuracy, with Area Under the Curve (AUC) values of 96.08% (MDD versus HC) and 91.77% (SCZ versus HC) in the main dataset. However, the MDD versus SCZ classification had an AUC of 57.75%. The motivation reward component, loss reward consequence, and high reward magnitude were the most influential features within respective categories for distinguishing both MDD and SCZ from HC (p < 0.001). A refined model using only the top eight features maintained robust performance, achieving AUCs of 96.06% (MDD versus HC) and 95.18% (SCZ versus HC).

Conclusion: The stacking model effectively classified SCZ and MDD from HC, contributing to understanding transdiagnostic mechanisms of anhedonia.

Keywords: anhedonia; machine learning; major depressive disorder; reward processing; schizophrenia; stacking model.

Figure 1.

The flow of the feature extraction. Note: (A) MID Paradigm Scheme. (a) Reward component: Anticipatory pleasure – Prediction (PRED); (c) Reward consequences: Gain/Loss (G/L); Reward magnitudes: High/Low/Control (H/L/C); (d) Reward component: Anticipatory pleasure- Feeling (FEEL); (e) Reward component: Consummatory pleasure (CONS); (f-b) Reward component: Motivation (MOTI); (g) Reward component: Remembered pleasure (Recall-RECA). (B)Combination of Feature Sets: The number within each circle represents the count of features included in that feature set. Detailed descriptions of the features are provided in Supplementary Table S3. GH, ‘gain high’; GL, ‘gain low’; GC, ‘gain control’; LH, ‘loss high’; LL, ‘loss low’; LC, ‘loss control’. (C)The workflow of Machine Learning.

Figure 2.

Importance of each feature set by stacking. Note: MDD, ‘major depressive disorder’; SCZ, ‘schizophrenia’; HC, ‘healthy controls’; MOTI, ‘motivation’; ANTI-FEEL, ‘feeling’; ANTI-PRED, ‘prediction’; CONS, ‘consummatory pleasure’; RECA, ‘remembered pleasure’; GAIN, ‘gain reward’; LOSS, ‘avoid loss reward’; GH, ‘gain high’; GL, ‘gain low’; GC, ‘gain control (control, no rewards)’; LH, ‘loss high’; LL, ‘loss low’; LC, ‘loss control’.