Abstract
Emotion dysregulation is a common issue across childhood and adolescence, including in autism and attention-deficit/hyperactivity disorder (ADHD). Yet most accounts remain descriptive—focusing on irritability, rapid escalation, slow return to baseline—without specifying the computational processes that stabilize emotions in context. We propose a predictive-coding account in which emotion regulation depends on how the brain predicts bodily and environmental signals, detects mismatches, between expectation and sensation, and determines how much confidence to assign those mismatches. Within this framework, dysregulation can arise when prediction errors are assigned excessive precision (i.e., incoming signals are overweighted relative to prior expectations, making them feel “too loud”), amplifying emotional reactivity or when prior expectations update too slowly (feelings get “stuck”), prolonging emotional states beyond contextual demands. We hypothesize that autism may involve heightened confidence assigned to surprising inputs together with sluggish updating of expectations, a combination that can produce both reactivity and emotional inertia. ADHD, by contrast, may be characterized by weaker expectations and unstable gain control, meaning the system’s amplification of incoming signals fluctuates from moment to moment, contributing to rapid, reactive swings in affect. We outline candidate physiological correlates—including heart-rate variability dynamics, pupillary variability, mismatch negativity, and neural indices of interoceptive precision—that could evaluate these hypotheses. Framing intervention around restoring flexible inference rather than suppressing emotion offers a mechanistic direction for developmental research. These predictions await empirical validation in developmental samples.
Emotion dysregulation (ED), difficulty flexibly modulating emotional responses to context and returning to baseline after perturbation, is increasingly recognized as a critical feature across neurodevelopmental disabilities, particularly autism and attention-deficit/hyperactivity disorder (ADHD) (Mazefsky et al., 2013; van Stralen, 2016). Despite its prevalence and impact, existing frameworks remain largely descriptive, emphasizing executive dysfunction or autonomic imbalance without specifying the underlying computational mechanisms of affective inference. We address this gap by proposing a predictive coding framework for ED across autism and ADHD. A schematic overview of this proposed framework is shown in Figure 1A. We argue that disruptions in hierarchical prediction, the brain’s assignment of confidence to incoming signals (precision-weighting), and updating of interoceptive and exteroceptive information can explain both the emotional volatility and the emotional inertia observed in these conditions. Co-occurrence of autism and ADHD (AuDHD) traits may further amplify these predictive instabilities, offering an informative test case for this unified model.
Predictive-coding framework for emotion dysregulation: Computational profiles and translational pathways.
Conceptually, this framework treats emotion regulation as a skill that develops through childhood and adolescence as predictive hierarchies become increasingly calibrated (Cole et al., 2004; Smith et al., 2017). Key developmental windows likely include early childhood (ages 3–7), when prefrontal–autonomic coupling begins to strengthen (Gatzke-Kopp and Beauchaine, 2007), middle childhood (ages 7–12), when interoceptive awareness becomes more differentiated and verbally accessible (Murphy et al., 2017), and adolescence (ages 12–18), when continued prefrontal maturation enables more sophisticated emotion regulation strategies (Smith et al., 2017). Deviations in predictive calibration during these windows—such as persistently low heart-rate variability (HRV) or failure to develop stable interoceptive awareness—may set the stage for later dysregulation. Longitudinal studies tracking precision-weighting and updating parameters across these developmental stages are needed to test when and how these processes go awry.
Predictive Coding: Core Computational Terms
Predictive coding posits that the brain operates as a hierarchical prediction machine, continually generating expectations about incoming inputs and minimizing prediction error by updating internal models or enacting corrective behavior (Clark, 2013; Friston, 2010). Within this architecture, emotions are inferred predictions about bodily and environmental states that guide adaptive regulation (Barrett and Simmons, 2015; Seth and Friston, 2016). Emotion regulation thus depends on how accurately and flexibly these predictive models are maintained and revised. In Bayesian terms, precision refers to the confidence assigned to a signal or belief—how much weight a prediction error is given relative to prior expectations (Friston, 2010). Gain control refers to neuromodulatory mechanisms, often catecholaminergic, that dynamically adjust this precision by amplifying or dampening incoming signals (Arnsten, 2009; Hauser et al., 2016). Volatility describes the estimated instability of the environment and determines how quickly priors are updated (Mathys et al., 2014). Throughout this article, we use these terms in their computational sense rather than as metaphors. In neurodevelopmental conditions, these predictive hierarchies may diverge from typical development.
In autism, evidence suggests slower updating of priors and overly high precision assigned to prediction error signals, leading to heightened sensory reactivity and cognitive rigidity (Lawson et al., 2014; Van de Cruys et al., 2014). In ADHD, by contrast, studies suggest underweighted priors and heightened volatility in precision allocation—that is, moment-to-moment fluctuations in how strongly the brain amplifies or dampens incoming sensory and interoceptive signals—producing distractibility and affective lability (Hauser et al., 2016; Hesp et al., 2021). Yet, while predictive-coding theories have been applied extensively to perception and attention, their implications for emotion regulation, and specifically for ED, remain underexplored.
This omission represents a key theoretical gap. Most developmental models of ED focus on trait-level constructs such as reactivity or executive control (Cole et al., 2004; Gross, 2015) and rarely specify how emotional states are dynamically inferred and stabilized. A predictive-coding account provides mechanistic precision by embedding emotion within hierarchical inference processes governed by prediction error and precision optimization—the ongoing adjustment of how much confidence the brain assigns to different incoming and internal signals. ED can be conceptualized as the consequence of imbalanced predictive hierarchies: When affective prediction errors are overweighted, emotions become volatile; when priors are overly weighted or update too slowly, emotions become inert and resistant to change (Owens et al., 2022; Smith et al., 2017).
This framework also connects with measurable physiological markers. HRV reflects autonomic flexibility and central–peripheral integration (Beauchaine, 2015; Thayer et al., 2012), while interoceptive inference models describe how precision-weighting of bodily signals shapes emotional awareness (Seth and Friston, 2016). These biomarkers bridge computational theory and developmental physiology, offering testable hypotheses for future research. By integrating predictive coding, interoceptive neuroscience, and developmental clinical science, our approach provides a unified, mechanistically grounded account of ED across autism and ADHD. It outlines a translational agenda for identifying measurable prediction-error dynamics as candidate biomarkers and potential intervention targets for emotional instability in neurodevelopmental populations.
Position relative to existing work
Prior applications of predictive coding to autism have focused on sensory prediction and the perception of others’ actions or emotions (Palmer et al., 2017; Van de Cruys et al., 2014), while developmental reviews address attention and cognition (Kilner et al., 2007). However, these studies stop short of modeling how predictive processes govern emotion regulation itself. Likewise, clinical reviews of ED in autism and ADHD (Mazefsky et al., 2013; Shaw et al., 2014) remain descriptive rather than mechanistic. Our approach advances this literature by centering emotion regulation within the predictive-coding framework, specifying key computational parameters—priors, prediction errors, precision, and updating speed—and linking them to measurable physiology and intervention targets. Interoceptive accounts of autism (Garfinkel et al., 2016; Owens et al., 2022) have touched on precision-weighting, but none have yet outlined how predictive processes might mechanistically shape emotion regulation. Here, we outline a predictive-coding framework specifying how predictive processes generate and regulate emotional states. The distinctions summarized in Table 1 provide the computational scaffolding for this framework.
Computational Parameters in Predictive Coding and Emotion Regulation
All biomarker mappings are proposed correlates rather than validated selective indices of computational parameters.
ACC, anterior cingulate cortex; EDA, electrodermal activity; HRV, heart-rate variability; MMN, mismatch negativity; PFC, prefrontal cortex.
Reality check: what this model does—and does not—claim
Direct tests of affective precision-weighting in developmental samples are still scarce. Most predictive-coding work in youth has focused on sensory or attentional processes rather than emotion regulation. Box 1 translates this framework into near-term, testable predictions for pediatric labs using accessible physiology and simple learning paradigms (Huys et al., 2021; Patzelt et al., 2018; Smith et al., 2020). We next outline the predictive-coding architecture linking emotional inference to precision-weighting and updating, which motivates Figure 1A. Figure 1B provides a simplified pathway from computational pathophysiology through measurable biomarkers to intervention targets, offering a translational bridge for clinicians. In practical terms, the model proposes a three-step chain: (1) specific computational imbalances (e.g., excessive precision assigned to affective prediction errors, slow updating of priors, or unstable gain control) give rise to distinct emotional phenotypes; (2) these imbalances are expected to manifest in measurable physiological patterns (e.g., altered HRV dynamics, exaggerated or volatile mismatch-negativity [MMN] responses, or unstable pupil-linked arousal); and (3) interventions may be selected based on which computational “lever” appears most dysregulated. Thus, the framework moves from abstract pathophysiology to observable physiology and then to targeted intervention hypotheses.
Critical caveat
The framework we present is largely hypothetical. While predictive-coding principles are well-established in perception and learning, their specific application to ED in autism and ADHD remains to be directly tested. The physiological measures we discuss (HRV, MMN, pupillometry) are established biomarkers of autonomic and neural function, but their interpretation as direct indices of Bayesian precision-weighting or updating speed remains a theoretical bridge requiring validation through studies that jointly measure these physiological signals and computationally estimated parameters from formal models.
Box 1. Falsifiable predictions derived from the predictive-coding model of emotion dysregulation
Acronyms: ADHD, attention-deficit/hyperactivity disorder; ED, emotion dysregulation; EDA, electrodermal activity; EEG, electroencephalography; HRV, heart-rate variability; IQ, intelligence quotient; MMN, mismatch negativity.
Note: Testable using frameworks from Beauchaine (2015), Craske et al. (2014), Lehrer and Gevirtz (2014), Owens et al. (2022), Mathys et al. (2014). Bolded headings indicate distinct, theoretically derived prediction domains rather than differential weighting or priority.
Predictive Coding as a Framework for Emotion Regulation
Predictive coding has become one of the most influential frameworks in neuroscience for explaining how the brain interprets and interacts with the world. The core idea is straightforward: Rather than passively responding to sensory input, the brain generates predictions from its internal model—its best estimate of what it expects to perceive—and compares these predictions to incoming signals. Mismatches, or prediction errors, are minimized through model updating or action (Clark, 2013; Friston, 2010). This iterative cycle of prediction and correction supports perception, learning, and adaptive behavior across multiple levels of the nervous system.
Emotions can be understood within this same predictive architecture. Instead of being reactions to external events, emotions are inferred states that summarize the brain’s best guess about the body’s internal condition and its relationship to the environment (Barrett and Simmons, 2015; Seth and Friston, 2016). For readers unfamiliar with predictive-coding terminology, Box 2 provides concise definitions of key computational terms used throughout this framework. From this perspective, the brain continuously predicts interoceptive signals, such as changes in heart rate, respiration, or gut activity, and adjusts them to maintain adaptive physiological stability, or allostasis (Sterling, 2012). When an unexpected bodily sensation arises, a mismatch between prediction and sensory inputs generates an affective prediction error. The brain then updates its internal model or alters physiology and behavior to reduce that error. In everyday life, this process allows a person to anticipate emotional outcomes, such as calming before public speaking or bracing for an unpleasant sound, and to regulate these states through experience-based expectation. Emotion regulation, then, reflects the capacity to flexibly update interoceptive and exteroceptive predictions in line with changing circumstances (Owens et al., 2022; Seth and Friston, 2016). Successful regulation requires balanced precision weighting—confidence in one’s prior expectations without ignoring new evidence. When the system assigns excessive precision to incoming interoceptive signals (up-weighting prediction errors), even small fluctuations in bodily state register as meaningful surprises, producing heightened reactivity. When priors dominate and incoming evidence is treated as unreliable, prediction errors are effectively down-weighted, resulting in emotional inertia or blunted responsiveness.
Box 2. Key terms inferred predictions
The brain’s best guesses about what sensory input (internal or external) will arrive next, based on past experience and current context.
These principles extend naturally to developmental disabilities, where differences in how the brain learns to predict and correct internal states may underlie persistent emotion-regulation difficulties. Recent advances in interoceptive neuroscience provide empirical support for this framework. Neuroimaging and psychophysiological studies link activity in the anterior insula and medial prefrontal cortex (PFC), regions central to interoceptive prediction and error signaling, to both emotional awareness and autonomic control (Barrett and Simmons, 2015; Smith et al., 2017). Variability in activity within these networks is associated with individual differences in emotional stability and resilience, reinforcing the idea that flexible prediction updating is key to adaptive affective regulation. This predictive view also offers one possible mechanism by which strategies that increase interoceptive accuracy or vagal flexibility, such as mindfulness or HRV biofeedback, may tend to improve emotion regulation (Beauchaine, 2015; Thayer et al., 2012).
In this way, predictive coding links computational inference to physiological regulation. By framing emotions as predictions about bodily and environmental states, it shifts the question from how strongly a person feels to how accurately and flexibly the brain learns to predict and adjust those feelings. This lens sets the stage for understanding why ED might emerge when predictive hierarchies become imbalanced, a theme developed in the following section.
Where Predictive Hierarchies Can Go Awry in Neurodevelopment
Emotion regulation depends on a calibrated balance between prior expectations and incoming sensory evidence. In neurodevelopmental conditions, this calibration may drift in different directions. Autism and ADHD both show disruptions in predictive balance, but we posit that they do so through distinct computational mechanisms. Each illustrates how the mechanisms that usually keep perception and emotion aligned can drift off course during development. In autism, two computational features—(1) overly precise prediction errors and (2) slow updating of priors—are related but distinct. Overprecision prediction makes discrepancies feel highly salient, contributing to hyperreactivity; slow updating means that it takes more evidence to revise internal expectations, contributing to emotional inertia. In ADHD, the pattern is different. Weak priors reflect uncertainty about what to expect, while unstable gain control—that is, moment-to-moment variability in how strongly incoming signals are amplified or suppressed—can yield affective lability, with rapid emotional shifts driven by fluctuating confidence in perceptual and interoceptive cues (Hauser et al., 2016).
Developmental calibration of prediction
Rather than redescribing predictive-coding principles, we now focus specifically on how these computational parameters may shift across development and how deviations at different stages could give rise to distinct trajectories of ED. Emotion regulation depends on systems that continue to mature through adolescence—particularly prefrontal–autonomic coupling and interoceptive awareness (Smith et al., 2017). From a predictive-coding standpoint, development refines how precisely the brain weights internal signals and how quickly it updates affective predictions.
These parameters do not mature simultaneously. Precision assigned to prediction-error signals may be relatively high in early childhood, when neuromodulatory systems that regulate gain are still stabilizing (Arnsten, 2009; Gatzke-Kopp and Beauchaine, 2007). As children accumulate experience, priors become more context-specific and increasingly constrained by learned regularities, particularly across middle childhood (McLaughlin et al., 2014). Updating speed, or learning rate, also appears to become more selective with age—rapid and broadly responsive early in development, then increasingly modulated by contextual reliability as cortical control systems mature (Patzelt et al., 2018). Gain control mechanisms supported by catecholaminergic systems and prefrontal circuitry continue to refine into adolescence (Arnsten, 2009; Smith et al., 2017). Deviations in these maturational trajectories—whether excessive precision, unstable gain, or persistently slow updating—may therefore produce distinct longitudinal patterns of ED.
Developmental hypotheses
We propose the following developmental benchmarks as testable hypotheses:
Early childhood (ages 3–7): Basic interoceptive differentiation emerges; children begin to distinguish different bodily states (hunger vs. fatigue) and link them to simple emotion labels. Prediction errors are highly weighted (high reactivity), and priors update rapidly. Expected trajectory: Increasing HRV and high trial-to-trial physiological variability.
Middle childhood (ages 7–12): Prefrontal–autonomic coupling strengthens; children develop more stable expectations about emotional contexts. Gain control begins to stabilize. Expected trajectory: HRV continues increasing; habituation to repeated stimuli becomes more evident; prediction updating becomes more contextually modulated rather than universally rapid.
Adolescence (ages 12–18): Continued prefrontal maturation enables hierarchical emotion regulation strategies; interoceptive precision becomes more finely tuned. Expected trajectory: Peak HRV is reached; updating speed becomes optimally flexible (neither too rigid nor too volatile); gain control stabilizes under most conditions.
Deviations
Youth with autism may show persistently slow habituation (ages 7+) and low HRV despite chronological maturation. Youth with ADHD may show persistently volatile gain control (high trial-to-trial variability) that fails to stabilize through adolescence. Co-occurring presentations may show mixed patterns—slow habituation combined with physiological volatility.
Autism: slow updating and overprecision
A growing body of work suggests that perception and learning in autism may be shaped by overly precise sensory prediction errors and inflexible updating of priors (Lawson et al., 2014; Palmer et al., 2017; Van de Cruys et al., 2014). In computational terms, incoming signals may be weighted as overly reliable relative to prior expectations. The result is a heightened sensitivity to change and a reduced ability to generalize from past experience. However, it is important to note that evidence for “overprecise interoceptive prediction errors” in autism remains mixed. While some studies report heightened interoceptive reactivity, others find intact (Schauder et al., 2015) or even reduced interoceptive accuracy (Garfinkel et al., 2016; Palser et al., 2021). The proposed mechanism may apply most strongly to specific interoceptive domains (e.g., cardiac vs. respiratory signals) or to subsets of autistic individuals rather than representing a universal feature.
Clinically, these dynamics can surface as insistence on sameness, intolerance of uncertainty, and difficulty adapting to emotional cues (South and Rodgers, 2017). When applied to affect, this model helps explain why some autistic individuals experience emotions that feel overwhelming or, on the other hand, “stuck.” If affective prediction errors are given too much weight, small shifts in internal signals can cascade into strong emotional reactions. Conversely, if priors are more heavily weighted, emotional states can persist long after the triggering event has passed. Both patterns, hyperreactivity and hyporeactivity or inertia, emerge naturally from disrupted precision balance. Although they are consistent with existing behavioral and physiological findings in autism and ADHD (Hauser et al., 2016; Lawson et al., 2014), direct studies that formally estimate precision parameters and compare them across diagnostic groups during affective tasks are still needed.
ADHD: noisy priors and volatile gain control
In ADHD, the imbalance is different. Rather than overly precise priors, evidence points to noisy or weak priors and unstable regulation of neural gain—that is, fluctuating network-level amplification of incoming sensory and interoceptive signals (Hauser et al., 2016; Hesp et al., 2021). These rapid shifts in how strongly the system weights new input can make prediction errors spike or drop unpredictably, disrupting the formation of stable internal models. This volatility aligns with core behavioral features of ADHD—distractibility, impulsivity, and affective lability (Shaw et al., 2014)—where attention and emotion can shift abruptly from one state to another.
It is important to note that ADHD itself shows considerable heterogeneity in arousal and regulatory patterns. Empirical work has identified multiple ADHD subprofiles with distinct physiological signatures, including high-arousal dysregulated presentations, low-arousal sluggish cognitive tempo profiles, and mixed presentations (Karalunas et al., 2014; Musser et al., 2011; Nigg et al., 2005). These subgroups may reflect different configurations of predictive imbalance—for instance, volatile gain control may be most pronounced in the high-arousal subtype, while weak priors may characterize sluggish presentations. Future work should test whether computational parameters map onto these empirically derived subtypes and whether interventions can be tailored accordingly. Such mapping would allow the framework to move beyond a unitary ADHD profile toward parameter-level stratification within diagnosis.
Emotionally, this dynamic may translate into quick surges of anger, excitement, or frustration followed by equally rapid dissipation. If the brain’s precision control over priors is too loose, small cues can hijack affective systems, while feedback loops that normally stabilize emotion fail to take hold. While gain control instability in ADHD is well-supported in perceptual and cognitive tasks (Hauser et al., 2016), direct evidence for this mechanism specifically in affective contexts remains limited. The extension from attentional gain to emotional gain represents a theoretical prediction requiring empirical testing. In developmental terms, this volatility may reflect immature or dysregulated catecholaminergic gain systems that influence both attention and emotion (Arnsten, 2009).
Shared and divergent predictive signatures
Although autism and ADHD differ in the direction of predictive imbalance, both reflect departures from optimal hierarchical calibration. Co-occurring traits (often colloquially referred to as “AuDHD”) may amplify both tendencies: slow prior updating from autism and fluctuating gain from ADHD, together heightening the mismatch between expectation and experience (Craig et al., 2015). This computational framing unites behavioral observations under a single principle: emotional instability arises when the brain’s confidence in its predictions no longer matches environmental signals. Whether through rigid priors or noisy precision control, the shared outcome is difficulty maintaining steady, context-appropriate emotional states.
Co-occurrence as a computational collision (AuDHD)
Between 30% and 50% of autistic youth also meet criteria for ADHD (Antshel et al., 2013; Leitner, 2014). Our framework hypothesizes that co-occurrence reflects computational interference between these mechanisms. By interference, we mean that the mechanisms do not merely sum; rather, unstable gain allocation may interact with slow prior updating in ways that qualitatively alter system dynamics. This contrasts with an additive model, in which autism-related and ADHD-related mechanisms would independently contribute to symptom severity without altering one another’s computational effects. Clinically, this may appear as rapid mood shifts within prolonged dysregulated episodes or difficulty disengaging from distress, even as the trigger changes. Youth with co-occurring traits show greater functional challenges than either diagnosis alone (Tye et al., 2014), consistent with compounded predictive dysfunction. Empirically, AuDHD profiles should show (a) high trial-to-trial variability in pupil/electrodermal activity responses (ADHD signature) combined with (b) slow habituation across blocks (autism signature). Because cognitive flexibility and set-shifting rely on distinct neural networks in autism and ADHD (Dajani and Uddin, 2015), intervention may need sequential targeting—first stabilizing gain (neurofeedback, predictable routines) to reduce noise, then accelerating updating (exposure, cognitive-flexibility training) to reduce inertia (Antshel & Russo, 2019). These proposals remain speculative and should be evaluated empirically by directly estimating precision, prior strength, and updating speed within co-occurring samples.
Boundary conditions and mixed findings
The predictive-coding account we propose must be evaluated against a literature showing substantial heterogeneity and, in some domains, mixed findings. Interoceptive processing in autism, for example, is far from uniform. While reviews describe atypical integration of bodily signals and altered neural tracking of interoceptive cues (DuBois et al., 2016), empirical studies report considerable variability: Some autistic individuals show reduced interoceptive accuracy, others show intact performance, and some demonstrate enhanced performance depending on task demands and measurement approach (Palser et al., 2021; Schauder et al., 2015). Differences across paradigms (heartbeat counting vs. tracking, neural indices vs. self-report confidence), developmental stage, and individual heterogeneity likely contribute to this variability. However, commonly used measures such as heartbeat counting have known limitations, including reliance on beliefs or estimation strategies rather than pure interoceptive sensitivity (Desmedt et al., 2022; Ring and Brener, 2018).
Similarly, MMN findings in autism vary by stimulus type and context. While reduced MMN to simple auditory deviants has been reported and interpreted as altered precision-weighting (Balsters et al., 2017). However, other studies show intact or even enhanced MMN responses under specific task conditions (Garrido et al., 2009). This suggests that predictive alterations may be domain-specific rather than global and may emerge most clearly under conditions of uncertainty, affective salience, or increased cognitive load.
Autonomic findings in ADHD also demonstrate heterogeneity. Although reduced HRV is commonly reported (Musser et al., 2011), arousal patterns differ across temperament-based subtypes, including high-arousal/dysregulated and low-arousal/sluggish cognitive tempo presentations (Karalunas et al., 2014; Nigg et al., 2005). These differences raise the possibility that predictive dysregulation may map onto distinct physiological profiles rather than a single ADHD signature.
Finally, some studies using neutral statistical-learning or perceptual paradigms report intact predictive processing in subsets of autistic youth and adults (Van de Cruys et al., 2018). These mixed findings underscore that predictive dysregulation is unlikely to be a universal or static trait. Instead, it may manifest selectively—by domain (interoceptive vs. exteroceptive), by context (neutral vs. affective), by developmental stage, or by subgroup. Identifying these boundary conditions will be critical for refining and empirically testing the model.
From Computation to Physiology: Measurable Correlates of Predictive Dysregulation
If emotion regulation reflects predictive inference, then disruptions in this process should leave physiological signatures. Translating computational parameters into measurable biological variables is essential for connecting theory to empirical data. Established biomarkers indexing autonomic flexibility, neural error signaling, and interoceptive processing offer candidate bridges between predictive-coding constructs and observable physiology.
Autonomic flexibility and HRV
HRV provides one of the clearest physiological indices of autonomic flexibility that may be relevant to predictive updating. HRV reflects the dynamic interplay between sympathetic and parasympathetic control over cardiac rhythms, with higher variability indicating greater adaptability to changing internal and external demands (Thayer et al., 2012). The neurovisceral-integration model links this flexibility to prefrontal–subcortical circuits involved in emotion regulation (Smith et al., 2017). In predictive-coding terms, HRV may reflect autonomic flexibility relevant to how the system adjusts bodily predictions in response to new feedback. Reduced HRV, frequently observed in individuals with emotional lability or anxiety, may be consistent with slower or less precise updating of autonomic predictions, though this interpretation remains inferential (Beauchaine, 2015). Importantly, HRV indexes autonomic flexibility more directly than specific Bayesian parameters; the link to precision-weighting and updating speed remains a theoretical interpretation requiring validation through computational modeling studies that jointly measure HRV and estimated learning rates to determine their true relationship to one another.
Important qualification
The biomarkers discussed here—HRV, MMN, pupil dilation—are multidetermined measures that reflect numerous physiological processes, not selective indices of specific Bayesian parameters. Low HRV, for example, can indicate reduced vagal tone, poor sleep quality, anxiety, physical deconditioning, or metabolic stress, not exclusively slow updating parameters (Thayer et al., 2012). Likewise, pupil dilation is influenced by attention, arousal, task engagement, and ambient lighting conditions alongside prediction-error magnitude (Liao et al., 2018). We therefore characterize these measures as candidate correlates—plausible but not definitive proxies—of computational constructs. Validation requires studies that simultaneously measure these physiological signals and estimate computational parameters through formal model fitting to demonstrate their true relationship. At present, no physiological measure can be considered a selective index of any single predictive parameter, and all proposed mappings should be interpreted as provisional hypotheses rather than established mechanistic correspondences.
Neural indices of prediction error
At the neural level, prediction-error signaling can be tracked through event-related potentials such as MMN. MMN reflects the brain’s automatic response to deviations from expected patterns and has been widely interpreted as an electrophysiological correlate of sensory prediction error (Garrido et al., 2009). Reduced or delayed MMN responses have been documented in both autism and ADHD and have been interpreted as reflecting altered sensory prediction error signaling (Balsters et al., 2017; Garrido et al., 2009).
Similar patterns appear in emotion-processing paradigms, where exaggerated neural responses to affective deviations may indicate altered sensitivity to unexpected emotional cues (Richey et al., 2015). Pupillometry provides a complementary peripheral measure of prediction error. Pupil dilation varies with noradrenergic arousal and has been associated with the processing of unexpected or salient events (Liao et al., 2018). Heightened or unstable pupil responses in autism and ADHD may be consistent with noisy or inconsistent precision control within affective processing systems.
Interoceptive accuracy and awareness
Because emotions are grounded in interoceptive inference, measures of interoceptive accuracy—how well individuals perceive internal bodily signals—offer another window into predictive regulation. Studies suggest that autism and ADHD populations may exhibit altered interoceptive awareness and reduced correspondence between physiological arousal and subjective feeling states (Garfinkel et al., 2016; Wiersema and Godefroid, 2018). These findings align with the view that emotional instability may stem from imprecise mappings between predicted and sensed internal states. Enhancing interoceptive accuracy through training or feedback may help recalibrate this mapping, reducing affective volatility.
Interoception as a precision target
Interoception offers a direct window on how the brain predicts and updates bodily state. Reviews in autism suggest atypical integration of internal and external cues, and—importantly for this article—differences in how reliably neural activity tracks bodily signals (DuBois et al., 2016). In predictive-coding terms, this is a precision question: How consistently does the system treat interoceptive evidence as reliable when updating affective expectations?
Operationally, interoceptive precision is often better approximated by neural reliability (e.g., trial-to-trial stability of heartbeat-evoked potentials or respiratory-linked electroencephalography [EEG]) than by heartbeat counting or confidence ratings, which can reflect strategy and metacognition as much as signal fidelity (Park & Blanke, 2019; Petzschner et al., 2017). In ADHD, by contrast, the more robust signature is often exteroceptive volatility—larger moment-to-moment fluctuations in responses to unexpected external events (e.g., variable MMN; unstable pupil-linked arousal). Put simply, we hypothesize that autism may show a stronger interoceptive-precision signature, whereas ADHD may show a stronger exteroceptive-volatility signature—distinct “failure points” within the same predictive hierarchy.
Integrating biomarkers with computational models
HRV, MMN, pupillary responses, and interoceptive accuracy form a convergent set of physiological readouts that provide a window into predictive dysregulation. Each captures a different level of the predictive hierarchy, from central error signaling to peripheral autonomic adjustment. Future studies that combine these measures could begin to quantify how individuals update internal models of emotion, providing a bridge between abstract computation and observable physiology. Such an approach could eventually help identify computationally informed subtypes of ED, pending empirical validation, those driven by overprecision versus those driven by sluggish updating, and tailor interventions accordingly. Operationalizing these predictions means designing tasks that can be practically used for children and adolescents. For instance, volatility-manipulated affect tasks could use simple emotional images or short film clips whose outcomes change in predictable versus unpredictable ways, allowing observation of how youth adjust to shifting emotional contingencies. Interoceptive precision can be measured with age-appropriate heartbeat-tracking tasks, where participants estimate their own heartbeat counts during fixed intervals while actual cardiac activity is measured. The discrepancy between perceived and actual cardiac signals may provide one behavioral proxy of interoceptive accuracy, though such tasks also reflect metacognitive and attentional factors. When these behavioral measures are paired with computational model fitting, such as the hierarchical Gaussian filter (Mathys et al., 2014), we can begin to estimate model-derived precision and updating parameters at the individual level. Box 1 summarizes concrete, falsifiable predictions derived from this framework, translating the proposed mechanisms into experimentally testable markers for child and adolescent samples.
Translational and Clinical Implications
Viewing ED through a predictive-coding lens shifts the goal of intervention. We are not just dampening big feelings; we are recalibrating inference, how the brain assigns confidence to predictions and updates them when the world (or the body) disagrees. This framework is not only explanatory but also clinically actionable. Understanding how predictive mechanisms mature—and where they may deviate—can guide early identification, prevention, and targeted intervention for children and adolescents showing early signs of dysregulation. If dysregulation reflects either (a) overprecise affective prediction errors or (b) slow, rigid priors, then interventions should aim to adjust precision and speed up adaptive updating. Dysregulation can also stem from unstable gain control, as seen in ADHD, where the system fluctuates in how strongly it weights incoming signals (Hauser et al., 2016; Hesp et al., 2021). In those cases, interventions that help stabilize gain—such as neurofeedback targeting attentional control or practices that reduce arousal volatility—may support more consistent emotional responding. Indeed, several existing tools already nudge these levers.
Autonomic and interoceptive training
HRV biofeedback trains vagal control via paced breathing, improving parasympathetic regulation and stress recovery (Lehrer and Gevirtz, 2014). Higher HRV is consistently linked to more flexible central–autonomic coupling, the physiological substrate of rapid, adaptive updating (Thayer et al., 2012). Mindfulness and body-awareness practice similarly strengthen interoceptive precision: Training sustained, nonreactive attention to bodily signals increases insula-mediated representation of interoceptive attention and supports more accurate mapping between feeling and physiology (Farb et al., 2013). Such approaches can help identify and correct misweighted affective signals before they spiral. While HRV biofeedback and mindfulness consistently improve emotional control, these studies do not measure the impact of these methods on precision-weighting or Bayesian updating directly. Their observed effects on physiological flexibility and error-driven learning are consistent with computational recalibration but remain hypotheses requiring formal testing (Browning et al., 2020; Goessl et al., 2017; Tang et al., 2015). Framing them in predictive-coding terms should therefore be read as a conceptual bridge, not as proof of mechanism.
Critical mechanistic caveat
While HRV biofeedback, mindfulness, and neurofeedback show reliable benefits for emotional control across multiple randomized trials (Goessl et al., 2017; Lehrer and Gevirtz, 2014; Tang et al., 2015; Wielgosz et al., 2019), the mechanisms underlying these effects remain uncertain. We do not yet have direct evidence that these interventions work by recalibrating Bayesian precision-weighting or accelerating prediction updating. The improvements could equally reflect enhanced attentional control, reduced baseline arousal, strengthened prefrontal inhibition, or placebo/expectancy effects. The predictive-coding interpretation offered here represents a mechanistic hypothesis that could guide future mediator analyses, for example, testing whether changes in model-estimated learning rates statistically mediate the effect of mindfulness on emotional outcomes. Until such studies are conducted, claims about specific computational mechanisms remain speculative.
Neurofeedback and precision recalibration
Neurofeedback offers another potential route to recalibrating precision. Real-time feedback from regulatory hubs (e.g., anterior insula, anterior cingulate cortex, medial PFC) can strengthen top-down control and teach the system how heavily to weigh incoming errors (Sitaram et al., 2017). Early evidence suggests that EEG-based neurofeedback targeting frontal rhythms can reduce emotional lability in ADHD and anxiety, although larger and more rigorous trials are needed (Enriquez-Geppert et al., 2017).
Sensory–attentional and contextual approaches
For many youth with autism, predictable rhythmic input and structured sensory contexts reduce the stream of “surprising” signals, which may serve to lower the background load of prediction error and stabilize affect (Hardy and LaGasse, 2013; South and Rodgers, 2017). Multisensory environments that give the child or adolescent control over intensity and timing appear especially helpful for regulation and engagement (Unwin et al., 2021).
Clinical Significance
This predictive-coding framework offers several practical takeaways for understanding and supporting youth with ED. (1) It helps explain why dysregulation often persists even after behavioral or skills-based therapies. If the underlying inference system—the brain’s way of predicting and interpreting bodily and emotional states—remains miscalibrated, the same challenges will resurface. Seeing ED as a difference in predictive learning rather than noncompliance or a failure of willpower can also help reduce stigma and change how clinicians and families talk about these experiences. (2) It points toward measurable, individualized intervention targets. For example, youth who show low HRV combined with high pupil-linked arousal may benefit most from neurofeedback aimed at stabilizing gain control, while those who recover slowly after stress might respond better to gradual exposure exercises that train the system to update predictions more flexibly. This can serve to move the field away from one-size-fits-all diagnostic labels toward mechanism-based personalization. (3) It reframes the goal of intervention itself—from suppressing emotion to helping the nervous system assign more appropriate weighting to its own signals.
Box 1 outlines concrete, testable predictions that researchers can begin examining using accessible physiological tools like HRV monitors, pupillometry, or EEG. Table 2 translates these computational ideas into observable clinical patterns and intervention strategies, with the potential to bridge theory and everyday practice. In doing so, this framework supports the broader shift toward transdiagnostic, mechanism-driven approaches in developmental mental health (Huys et al., 2016). Table 2 aligns the computational parameters described above with observable emotion-regulation profiles, associated biomarkers, and potential intervention targets. Importantly, we distinguish between conceptual computational “targets” (e.g., excessive precision assigned to prediction-error signals, slow updating of priors, unstable gain control) and currently available clinical tools that may influence these processes indirectly. Table 2, therefore, specifies both the hypothesized parameter being modified and the practical feasibility of implementing each intervention in pediatric settings. Importantly, these mappings are not modality-specific; the same computational parameter can be probed using multiple neural and physiological measures, each capturing different aspects of the underlying process.
Computational Parameters in Predictive Coding and Emotion Regulation
Neural systems and measurement approaches are illustrative and not specific to individual computational parameters. All listed measures are multidetermined and should be interpreted as candidate correlates rather than selective indices. Intervention targets are hypothesis-driven and not established mechanisms.
ACC, anterior cingulate cortex; EDA, electrodermal activity; EEG, electroencephalography; fMRI, functional magnetic resonance imaging; HRV, heart-rate variability; MMN, mismatch negativity; PFC, prefrontal cortex.
Conclusion
ED cuts across diagnostic boundaries, but its underlying mechanisms have remained elusive. We propose that a predictive-coding framework offers a unifying way to understand these mechanisms across autism and ADHD. By viewing emotion as inference about bodily and environmental states, the model posits why emotional instability can arise when predictive hierarchies are altered, whether through overly precise high precision assigned to prediction error signals or sluggish updating of priors. In such a unifying framework, the same computations that shape perception and learning also govern how the brain anticipates and regulates feeling.
Future research needs to translate these ideas into quantifiable features. Computational modeling studies could illustrate how prediction-error dynamics differ across autism, ADHD, and combined profiles. Longitudinal work could track how these differences emerge over development, and how interventions, such as HRV biofeedback, neurofeedback, or interoceptive training, modify predictive updating over time. Integrating physiological markers like HRV, MMN, and pupillometry will allow direct tests of whether emotion regulation can indeed be reframed as a problem of prediction rather than one of pure behavioral control. Clinically, this approach broadens how we think about intervention. Instead of targeting surface behaviors, it encourages recalibration of the systems that generate emotional predictions in the first place. An individual whose nervous system overestimates threat may benefit from interventions that reduce precision on interoceptive errors; another who remains “stuck” in a prior emotional state might need training that accelerates updating. Such distinctions can help move the field from broad diagnostic categories toward mechanism-based personalization.
A key next step is to formalize this personalization approach by quantifying individual-level computational parameters within (not just across) diagnostic groups. Autism and ADHD populations are far from homogeneous: Interoceptive ability and emotional reactivity vary widely in autism (Palser et al., 2021; Schauder et al., 2015), and autonomic profiles differ across ADHD temperament subtypes (Musser et al., 2011). Computational phenotyping—estimating precision and updating parameters at the individual level—could identify subgroups that respond differently to intervention (Huys et al., 2016). For example, youth showing low HRV with high pupil volatility may benefit from neurofeedback to stabilize gain control, whereas those with rigid expectations and slow habituation may respond better to updating-focused exposure exercises. This direction aligns with person-centered approaches in developmental clinical science (Karalunas et al., 2014; Lombardo et al., 2019).
At a broader level, applying predictive-coding theory to ED aligns with current efforts in computational psychiatry to move beyond symptom clusters and identify shared neural computations across conditions (Huys et al., 2016). It invites collaboration among neuroscientists, clinicians, and computational modelers to connect theory with practice. As predictive frameworks become more refined, they may ultimately guide individualized approaches to emotion regulation, helping transform how we understand and support young people whose emotions often feel too much, or not enough, for the world around them.
Disclosures
All authors declare that they have no financial or institutional relationships with commercial entities (including pharmaceutical companies) that could be perceived as a potential conflict of interest. The authors declare no conflicts of interest.