Explainable AI enables clinical trial patient selection to retrospectively improve treatment effects in schizophrenia.

BACKGROUND: Heterogeneity among patients' responses to treatment is prevalent in psychiatric disorders. Personalized medicine approaches-which involve parsing patients into subgroups better indicated for a particular treatment-could therefore improve patient outcomes and serve as a powerful tool in patient selection within clinical trials. Machine learning approaches can identify patient subgroups but are often not "explainable" due to the use of complex algorithms that do not mirror clinicians' natural decision-making processes. METHODS: Here we combine two analytical approaches-Personalized Advantage Index and Bayesian Rule Lists-to identify paliperidone-indicated schizophrenia patients in a way that emphasizes model explainability. We apply these approaches retrospectively to randomized, placebo-controlled clinical trial data to identify a paliperidone-indicated subgroup of schizophrenia patients who demonstrate a larger treatment effect (outcome on treatment superior than on placebo) than that of the full randomized sample as assessed with Cohen's d. For this study, the outcome corresponded to a reduction in the Positive and Negative Syndrome Scale (PANSS) total score which measures positive (e.g., hallucinations, delusions), negative (e.g., blunted affect, emotional withdrawal), and general psychopathological (e.g., disturbance of volition, uncooperativeness) symptoms in schizophrenia. RESULTS: Using our combined explainable AI approach to identify a subgroup more responsive to paliperidone than placebo, the treatment effect increased significantly over that of the full sample (p < 0.0001 for a one-sample t-test comparing the full sample Cohen's d = 0.82 and a generated distribution of subgroup Cohen's d's with mean d = 1.22, std d = 0.09). In addition, our modeling approach produces simple logical statements (if-then-else), termed a "rule list", to ease interpretability for clinicians. A majority of the rule lists generated from cross-validation found two general psychopathology symptoms, disturbance of volition and uncooperativeness, to predict membership in the paliperidone-indicated subgroup. CONCLUSIONS: These results help to technically validate our explainable AI approach to patient selection for a clinical trial by identifying a subgroup with an improved treatment effect. With these data, the explainable rule lists also suggest that paliperidone may provide an improved therapeutic benefit for the treatment of schizophrenia patients with either of the symptoms of high disturbance of volition or high uncooperativeness. TRIAL REGISTRATION: clincialtrials.gov identifier: NCT 00,083,668; prospectively registered May 28, 2004.

Machine Learning Models Identify Multimodal Measurements Highly Predictive of Transdiagnostic Symptom Severity for Mood, Anhedonia, and Anxiety.

BACKGROUND: Insights from neuroimaging-based biomarker research have not yet translated into clinical practice. This translational gap may stem from a focus on diagnostic classification, rather than on prediction of transdiagnostic psychiatric symptom severity. Currently, no transdiagnostic, multimodal predictive models of symptom severity that include neurobiological characteristics have emerged. METHODS: We built predictive models of 3 common symptoms in psychiatric disorders (dysregulated mood, anhedonia, and anxiety) from the Consortium for Neuropsychiatric Phenomics dataset (N = 272), which includes clinical scale assessments, resting-state functional magnetic resonance imaging (MRI), and structural MRI measures from patients with schizophrenia, bipolar disorder, and attention-deficit/hyperactivity disorder and healthy control subjects. We used an efficient, data-driven feature selection approach to identify the most predictive features from these high-dimensional data. RESULTS: This approach optimized modeling and explained 65% to 90% of variance across the 3 symptom domains, compared to 22% without using the feature selection approach. The top performing multimodal models retained a high level of interpretability that enabled several clinical and scientific insights. First, to our surprise, structural features did not substantially contribute to the predictive strength of these models. Second, the Temperament and Character Inventory scale emerged as a highly important predictor of symptom variation across diagnoses. Third, predictive resting-state functional MRI connectivity features were widely distributed across many intrinsic resting-state networks. CONCLUSIONS: Combining resting-state functional MRI with select questions from clinical scales enabled high prediction of symptom severity across diagnostically distinct patient groups and revealed that connectivity measures beyond a few intrinsic resting-state networks may carry relevant information for symptom severity.

Machine Learning Identifies Large-Scale Reward-Related Activity Modulated by Dopaminergic Enhancement in Major Depression.

BACKGROUND: Theoretical models have emphasized systems-level abnormalities in major depressive disorder (MDD). For unbiased yet rigorous evaluations of pathophysiological mechanisms underlying MDD, it is critically important to develop data-driven approaches that harness whole-brain data to classify MDD and evaluate possible normalizing effects of targeted interventions. Here, using an experimental therapeutics approach coupled with machine learning, we investigated the effect of a pharmacological challenge aiming to enhance dopaminergic signaling on whole-brain response to reward-related stimuli in MDD. METHODS: Using a double-blind, placebo-controlled design, we analyzed functional magnetic resonance imaging data from 31 unmedicated MDD participants receiving a single dose of 50 mg amisulpride (MDDAmisulpride), 26 MDD participants receiving placebo (MDDPlacebo), and 28 healthy control subjects receiving placebo (HCPlacebo) recruited through two independent studies. An importance-guided machine learning technique for model selection was used on whole-brain functional magnetic resonance imaging data probing reward anticipation and consumption to identify features linked to MDD (MDDPlacebo vs. HCPlacebo) and dopaminergic enhancement (MDDAmisulpride vs. MDDPlacebo). RESULTS: Highly predictive classification models emerged that distinguished MDDPlacebo from HCPlacebo (area under the curve = 0.87) and MDDPlacebo from MDDAmisulpride (area under the curve = 0.89). Although reward-related striatal activation and connectivity were among the most predictive features, the best truncated models based on whole-brain features were significantly better relative to models trained using striatal features only. CONCLUSIONS: Results indicate that in MDD, enhanced dopaminergic signaling restores abnormal activation and connectivity in a widespread network of regions. These findings provide new insights into the pathophysiology of MDD and pharmacological mechanism of antidepressants at the system level in addressing reward processing deficits among depressed individuals.

Intrinsic frequency biases and profiles across human cortex.

Recent findings in monkeys suggest that intrinsic periodic spiking activity in selective cortical areas occurs at timescales that follow a sensory or lower order-to-higher order processing hierarchy (Murray JD, Bernacchia A, Freedman DJ, Romo R, Wallis JD, Cai X, Padoa-Schioppa C, Pasternak T, Seo H, Lee D, Wang XJ. Nat Neurosci 17: 1661-1663, 2014). It has not yet been fully explored if a similar timescale hierarchy is present in humans. Additionally, these measures in the monkey studies have not addressed findings that rhythmic activity within a brain area can occur at multiple frequencies. In this study we investigate in humans if regions may be biased toward particular frequencies of intrinsic activity and if a full cortical mapping still reveals an organization that follows this hierarchy. We examined the spectral power in multiple frequency bands (0.5-150 Hz) from task-independent data using magnetoencephalography (MEG). We compared standardized power across bands to find regional frequency biases. Our results demonstrate a mix of lower and higher frequency biases across sensory and higher order regions. Thus they suggest a more complex cortical organization that does not simply follow this hierarchy. Additionally, some regions do not display a bias for a single band, and a data-driven clustering analysis reveals a regional organization with high standardized power in multiple bands. Specifically, theta and beta are both high in dorsal frontal cortex, whereas delta and gamma are high in ventral frontal cortex and temporal cortex. Occipital and parietal regions are biased more narrowly toward alpha power, and ventral temporal lobe displays specific biases toward gamma. Thus intrinsic rhythmic neural activity displays a regional organization but one that is not necessarily hierarchical.NEW & NOTEWORTHY The organization of rhythmic neural activity is not well understood. Whereas it has been postulated that rhythms are organized in a hierarchical manner across brain regions, our novel analysis allows comparison of full cortical maps across different frequency bands, which demonstrate that the rhythmic organization is more complex. Additionally, data-driven methods show that rhythms of multiple frequencies or timescales occur within a particular region and that this nonhierarchical organization is widespread.

Sentence processing in anterior superior temporal cortex shows a social-emotional bias.

The anterior region of the left superior temporal gyrus/superior temporal sulcus (aSTG/STS) has been implicated in two very different cognitive functions: sentence processing and social-emotional processing. However, the vast majority of the sentence stimuli in previous reports have been of a social or social-emotional nature suggesting that sentence processing may be confounded with semantic content. To evaluate this possibility we had subjects read word lists that differed in phrase/constituent size (single words, 3-word phrases, 6-word sentences) and semantic content (social-emotional, social, and inanimate objects) while scanned in a 7T environment. This allowed us to investigate if the aSTG/STS responded to increasing constituent structure (with increased activity as a function of constituent size) with or without regard to a specific domain of concepts, i.e., social and/or social-emotional content. Activity in the left aSTG/STS was found to increase with constituent size. This region was also modulated by content, however, such that social-emotional concepts were preferred over social and object stimuli. Reading also induced content type effects in domain-specific semantic regions. Those preferring social-emotional content included aSTG/STS, inferior frontal gyrus, posterior STS, lateral fusiform, ventromedial prefrontal cortex, and amygdala, regions included in the "social brain", while those preferring object content included parahippocampal gyrus, retrosplenial cortex, and caudate, regions involved in object processing. These results suggest that semantic content affects higher-level linguistic processing and should be taken into account in future studies.

Gamma- and theta-band synchronization during semantic priming reflect local and long-range lexical-semantic networks.

Anterior and posterior brain areas are involved in the storage and retrieval of semantic representations, but it is not known how these areas dynamically interact during semantic processing. We hypothesized that long-range theta-band coherence would reflect coupling of these areas and examined the oscillatory dynamics of lexical-semantic processing using a semantic priming paradigm with a delayed letter-search task while recording subjects' EEG. Time-frequency analysis revealed facilitation of semantic processing for Related compared to Unrelated conditions, which resulted in a reduced N400 and reduced gamma power from 150 to 450ms. Moreover, we observed greater anterior-posterior theta coherence for Unrelated compared to Related conditions over the time windows 150-425ms and 600-900ms. We suggest that while gamma power reflects activation of local functional networks supporting semantic representations, theta coherence indicates dynamic coupling of anterior and posterior areas for retrieval and post-retrieval processing and possibly an interaction between semantic relatedness and working memory.

Word class and context affect alpha-band oscillatory dynamics in an older population.

Differences in the oscillatory EEG dynamics of reading open class (OC) and closed class (CC) words have previously been found (Bastiaansen et al., 2005) and are thought to reflect differences in lexical-semantic content between these word classes. In particular, the theta-band (4-7 Hz) seems to play a prominent role in lexical-semantic retrieval. We tested whether this theta effect is robust in an older population of subjects. Additionally, we examined how the context of a word can modulate the oscillatory dynamics underlying retrieval for the two different classes of words. Older participants (mean age 55) read words presented in either syntactically correct sentences or in a scrambled order ("scrambled sentence") while their EEG was recorded. We performed time-frequency analysis to examine how power varied based on the context or class of the word. We observed larger power decreases in the alpha (8-12 Hz) band between 200-700 ms for the OC compared to CC words, but this was true only for the scrambled sentence context. We did not observe differences in theta power between these conditions. Context exerted an effect on the alpha and low beta (13-18 Hz) bands between 0 and 700 ms. These results suggest that the previously observed word class effects on theta power changes in a younger participant sample do not seem to be a robust effect in this older population. Though this is an indirect comparison between studies, it may suggest the existence of aging effects on word retrieval dynamics for different populations. Additionally, the interaction between word class and context suggests that word retrieval mechanisms interact with sentence-level comprehension mechanisms in the alpha-band.