Corrigendum to “On the integrity of functional brain networks in schizophrenia, Parkinson's disease, and advanced age: Evidence from connectivity‐based single‐subject classification.”

In our recent paper, subject‐wise resting‐state functional connectivity (RSFC) within each brain network was determined by extracting the time series for all nodes based on 6‐mm spheres around the meta‐analytically derived peaks. Subsequently, edge‐wise RSFC between all nodes was computed as pairwise Pearson correlations between all the time series and served as input features to the support vector machine (SVM) classifications. Unfortunately, we now discovered an error in the implementation of the time‐series extraction due to a typo in the respective code. In particular, we extracted all relevant voxels from the preprocessed resting‐state time series in one step and then split them according to the nodes they belonged to based on a linear index. When constructing the voxel list and the index based on the distance of all voxels from the respective meta‐analytical peaks, we used at one point “less than” and the other “less or equal”, which for a subset of nodes resulted in a frameshift causing a few voxels to be assigned to the wrong node. Having recognized this issue, we re‐ran all network extractions and classifications as specified in [Pläschke et al., 2017]. Importantly, results did not change, as we expected due to only a small subset of voxels being miss‐assigned. In the following, we update all findings with the correct version, providing an adjusted Table 1 and Figure 2 as well as a revision of the Supporting Information.

Table 1.

Classification results of the support vector machine of all groups based on specific networks

Network (Abbr.)	SCZ vs. HCSCZ Acc. (Sens. / Spec.) AUC	PD vs. HCPD Acc. (Sens. / Spec.) AUC	Old vs. Young Acc. (Sens. / Spec.) AUC
EmoSF	70% (78% / 62%) 0.78	65% (65% / 64%) 0.71	87% (88% / 85%) 0.95
ER	70% (80% / 60%) 0.77	64% (67% / 62%) 0.70	80% (82% / 78%) 0.87
ToM	65% (76% / 53%) 0.70	68% (74% / 63%) 0.74	80% (80% / 79%) 0.88
Empathy	71% (71% / 71%) 0.80	65% (66% / 64%) 0.74	78% (78% / 78%) 0.86
Rew	74% (74% / 74%) 0.80	65% (64% / 66%) 0.75	91% (91% / 92%) 0.96
AM	64% (70% / 59%) 0.72	72% (74% / 70%) 0.77	82% (83% / 81%) 0.90
SM	71% (76% / 66%) 0.77	70% (67% / 72%) 0.81	85% (86% / 83%) 0.92
WM	68% (69% / 67%) 0.77	68% (67% / 70%) 0.75	83% (83% / 83%) 0.90
CogAC	64% (70% / 58%) 0.67	66% (64% / 69%) 0.75	81% (82% / 80%) 0.92
VigAtt	69% (73% / 64%) 0.73	69% (69% / 68%) 0.74	84% (84% / 85%) 0.92
MNS	56% (65% / 46%) 0.56 #	53% (53% / 52%) 0.46 #	82% (83% / 81%) 0.88
Motor	58% (69% / 47%) 0.51 #	68% (66% / 70%) 0.77	75% (76% / 75%) 0.87

Abbreviations: Acc., Accuracy (in %)/Sens., sensitivity (in %)/Spec., specificity (in %)/AUC, area under the ROC curve;

EmoSF, emotional scene/ face processing, ER, cognitive emotion regulation, ToM, theory‐of‐mind cognition, Empathy, empathic processing, Rew, reward‐related decision making, AM, autobiographical memory, SM, semantic memory, WM, working memory, CogAC, cognitive action control, VigAtt, vigilant attention, MNS, mirror neuron system, Motor, motor execution.

^# Network with no significant classification result.

Acc. refers to the proportion of subjects correctly classified as patients (PD, SCZ) or older age and subjects correctly classified as being healthy or younger age (mean of sensitivity and specificity). Sensitivity relates to the percentage of patients (SCZ or PD) correctly classified as being ill or else subjects correctly identified as old in the aging sample (true positives). Specificity relates to the percentage of healthy subjects correctly classified as being healthy or else subjects correctly identified as young in the aging sample (true negatives). AUC refers to the area under the ROCs curve. The ROC curve depicts the relationship between true positive rate and false positive rate.

Figure 2.

Group classification results of the SVM. (a) Polar plot of group classification accuracies based on all 12 networks for SCZ (in green), PD (in blue) and NA (in yellow). Accuracy refers to the proportion of subjects correctly classified as patients (PD, SCZ) or older age and subjects correctly classified as being HCs or younger age. (b) Polar plot of z‐standardized accuracies (corrected for multiple comparisons) of patients classification for SCZ (in green) and PD (in blue). (c) Log‐likelihood ratios of classification performance for networks showing higher classification for one patient group vs. the other

1. RESULTS

For schizophrenia (SCZ), networks related to reward processing (accuracy [Acc.] = 74%; area under the receiver‐operating curve [AUC] = 0.8), empathy (Acc. = 71%; AUC = 0.8), cognitive emotion regulation (Acc. = 70%; AUC = 0.77), as well as emotional processing (Acc. = 70%; AUC = 0.78) distinguished patients most accurately from their healthy controls (HCs). The reward processing network was significantly better in the SCZ classification compared with all other networks (p < 0.001). For Parkinson's disease (PD), the networks subserving autobiographical memory (Acc. = 72%; AUC = 0.77), motor execution (Acc. = 68%; AUC = 0.77), and theory‐of‐mind cognition (Acc. = 68%; AUC = 0.74) yielded the highest classification accuracies. The autobiographical memory network was significantly better in the PD classification compared with all other networks (p < 0.001). All network comparison results within the patient groups are summarized in Table SV and SVI.

The comparison of classification performance between SCZ and PD confirmed that the networks discriminating either disorder from their respective controls were specific to that particular disorder (Figure 2B,C). Networks which were most accurate in distinguishing SCZ from HCs (reward processing, empathy, cognitive emotion regulation and emotional processing) exhibited significantly better classification performance in SCZ vs. HC classification compared with the PD vs. HC classification (p < 0.001; Table SVII). Likewise, networks which performed best at discriminating PD patients from HCs (autobiographical memory, motor execution and theory‐of‐mind cognition) showed significantly better classification performance compared with the SCZ vs. HC classification (p < 0.001; Table SVII).

As already noted in the original publication, each single network was better at discriminating young from old participants than any network was at classifying either PD or SCZ patients (Table SIX and SX). While in the young vs. old classification each network still yielded accuracies well above chance (≥ 75%), we now note, though, that the reward processing network not only yielded the highest accuracy (Acc. = 91%; AUC = 0.96) but was also significantly better at classifying young vs. old subjects than any of the other networks (p < 0.001; Table SVIII).

2. DISCUSSION

2.1. Classification of Schizophrenia Patients versus Controls

Both the emotional processing and empathy network are still highly specific networks for SCZ classification, even though the emotional processing network did not achieve the highest accuracy any more after correcting the data extraction error. Also, we found a decrease in the performance of the network subserving cognitive action control for SCZ classification (Acc. = 64%; AUC = 0.67). While this network still showed a classification performance for SCZ vs. HC that was significantly above chance, the decrease in accuracy combined with an increase in PD‐classification accuracy now renders this network similarly discriminative for both disorders, which actually matches the clinical deficits in cognitive action control in patients with either PD and SCZ [Lesh et al., 2011; Manza et al., 2017]. Interestingly, the re‐analysis now shows the highest accuracy with respect to SCZ vs. HC classification for the reward processing network. This actually matches our proof‐of‐concept hypothesis for SCZ, given the well‐established neural alterations associated with reward‐learning in this disorder, while the previous, erroneous, observation that this network did not differentiate SCZ from HCs with higher accuracy represented somewhat of a surprise. The corrected result now corroborates the prominent role of the dysfunctional reward system associated with dopaminergic alteration in SCZ [Toda and Abi‐Dargham, 2007] and aberrant salience processing in psychosis [Heinz and Schlagenhauf, 2010; Radua et al., 2015]. Moreover, the cognitive emotion regulation network now comes out as more specific for SCZ classification, which is, however, not attributable to a significant increase in SCZ classification accuracy, but rather to a decrease in classification accuracy for PD. Nevertheless, this is in line with dysfunctional emotion regulation as a well‐established characteristic of SCZ [Khoury and Lecomte, 2012; van der Meer et al., 2014] and is well mirrored in the degree of SCZ‐related information that is contained in the cognitive emotion regulation network.

2.2. Classification of PD Patients versus Controls

We found that the network subserving semantic memory, originally reported to show one of the most specific classification performances for PD, has lost this feature due to an increase in its classification performance for SCZ vs. HC. Of note, the accuracy for its PD vs. HC classification has remained more or less unchanged and is still significantly above chance (Acc. = 70%; AUC = 0.81). Hence, we would argue that the semantic memory network contains relevant information on PD pathophysiology, even though it now appears not to be specific to this disorder as relevant information on SCZ are likewise contained. All other networks that initially yielded superior classification performance, i.e., those related to autobiographic memory, motor functions, and theory‐of‐mind cognition, remained highly specific for PD.

2.3. Age Group Classification

The reward processing network showed outstanding performance in the young vs. old classification. The fact that this network even outperformed the SCZ and PD classifications indicates the relevance of age‐related changes associated with the reward system [Vink et al., 2015] as a marker for age group classification.

3. Conclusion

Re‐analysis after adjustment of the time series extraction in principle replicated the originally reported results, with only slight changes in the relative performance for a few individual networks. In particular, the corrected results corroborate the two main conclusions stated in the original publication. First, both SCZ and PD are well predicted by distinct networks that resonate well with known clinical and pathophysiological features. Second, all networks yield accuracies for the adult age group classification that outperform any clinical classification accuracies.

Supporting information

Appendix S1: Supporting Information