People with psychotic spectrum disorders who have a documented history of interpersonal aggression make a specific, quantifiable cognitive error at roughly twice the rate of healthy individuals: they misread negative emotions — anger, fear, disgust, sadness decoded as something friendly or benign. A large, carefully controlled Swedish study has now isolated and measured that error type for the first time across three sensory modalities simultaneously, identifying it as a distinct neurocognitive mechanism that may disable the social feedback loop that normally stops confrontations before they escalate.

The peer-reviewed findings, published June 24, 2026, in Schizophrenia, a journal in Springer Nature’s Nature Portfolio, come from a multimodal study at Karolinska Institutet led by Gabriela Gavalova of Karolinska Institutet’s Department of Clinical Neuroscience and published as part of the Stockholm’s Project for Forensic Psychiatry (SPRätt). The research points toward a clinical implication that current forensic care largely overlooks: a standardized emotion recognition test already exists that adds measurable and independent predictive value over and above the structured professional judgment tools that forensic risk tools rely on today. If that finding holds up in follow-on longitudinal studies, it would give clinicians a reproducible cognitive dimension of violence risk assessment that does not depend on a patient’s self-report or prior record alone.

The study’s negative-to-positive misclassification rate — 14% in the aggression group versus roughly 7% in healthy controls — reached statistical significance across all 12 emotions and all three tested sensory channels, suggesting the error is not modality-specific or emotion-specific but reflects a broad-based perceptual and cognitive deficit — confusion matrix results X² significance at p

How Multimodal Emotion Testing Catches What Single-Channel Tests Miss

The study’s core assessment instrument, the Emotion Recognition Assessment in Multiple Modalities (ERAM) test, was developed at Karolinska Institutet specifically for the SPRätt project and represents a methodological step beyond the traditional six-emotion, static-image paradigm that has dominated social cognition research for decades — see ERAM test methods section. Instead of showing participants still photographs of actors’ faces and asking them to label the depicted emotion, ERAM presents 72 unique dynamic stimuli — short video clips and audio recordings of trained professional actors performing nonsense-word sentences to eliminate any linguistic cues — and asks participants to identify one of 12 possible emotions based on the twelve-emotion ERAM stimulus design: anger, anxiety, contempt, despair, disgust, embarrassment, fear, happiness, neutral affect, pain, sadness, or surprise.

Crucially, every stimulus is presented three ways: video only (face and body visible, no sound), audio only (voice audible, screen blank), and synchronized audiovisual (both channels simultaneously). This three sensory modality design allows researchers to test not just whether a participant can recognize an emotion, but whether their deficit is specific to one sensory channel — a perceptual impairment — or reflects a more systemic cognitive failure that persists regardless of how the information is delivered.

The answer, across all three participant groups, was that multimodal presentation consistently produced the best recognition scores — per multimodal recognition Hu results, mean unbiased hit rate Hu = 0.37, outperforming video-only (Hu = 0.27) and audio-only (Hu = 0.25) conditions. The unbiased hit rate (Hu) metric is a statistical correction for forced-choice response bias — it adjusts raw accuracy scores to account for the fact that participants might guess the same answer repeatedly under uncertainty, which would inflate naive accuracy counts. By using Hu rather than raw percentage correct, the ERAM results reflect genuine recognition rather than strategic response patterns.

The study also generated confusion matrices — mathematical grids showing not just how often participants were wrong, but specifically what they guessed instead of the correct answer. Those matrices are where the negative-to-positive misclassification finding emerges: a systematic, group-level pattern of calling anger, fear, or sadness something positive rather than an equally negative or neutral emotion. That directional specificity — negative decoded as positive — is what gives the finding its clinical weight, because it describes the precise cognitive distortion that could cause someone to walk toward a danger signal rather than away from it.

Three Groups, 237 Participants, and a Strict Standard for Violence

The study drew from a cohort assembled through the SPRätt project at Stockholm University, the Stockholm’s Project for Forensic Psychiatry, a multi-institution research program that ran from 2018 through 2024 at Karolinska Institutet, Uppsala University, and Stockholm University.

Three carefully matched groups participated. The first group — 79 individuals with psychotic spectrum disorders and a documented history of interpersonal aggression — was recruited from Forensic Psychiatry Stockholm per aggression group participant recruitment. The second — 72 people with psychotic spectrum disorders but no aggressive history — was recruited from general psychiatric outpatient clinics across Stockholm. A control group of 86 healthy individuals, recruited through Sweden’s State Person Address Registry, completed the sample. Groups were matched on age and gender demographics; men made up between 69% and 75% of each group, with mean ages ranging from 35 to 39 years.

Psychotic spectrum disorders in the study encompassed schizophrenia, schizoaffective disorder, delusional disorder, and bipolar disorder with psychotic features. Schizophrenia dominant diagnosis rates were striking — 79% of the aggression group and 86% of the non-aggression group. All participants with psychiatric diagnoses were clinically stable, with no medication changes for at least three months before testing.

The threshold for classifying a participant as having a history of aggression was deliberately stringent based on aggression classification criteria: legal-level acts only — threats to life, assault, grievous bodily harm, or murder and manslaughter — verified through crime records, forensic psychiatric assessments, hospital case files, and self-reports. About half of the aggression group had committed assault; roughly a third had been convicted of grievous bodily harm.

Distinctive Error, Not Just Worse Scores

All three groups performed well above chance on the ERAM, but ERAM accuracy results table showed substantial and consistent gaps. Participants with psychotic spectrum disorders and a history of aggression posted the weakest accuracy across all 12 emotions, ranging from 25% correct for despair to 54% for anger. Non-aggressive psychosis patients performed somewhat better (29% to 66%), while healthy controls performed best (39% to 68%). Across all groups, anger was the most recognizable emotion; anxiety and despair the hardest to identify.

The consistency of these gaps across all 12 emotions and all three modalities is itself a significant finding. Rather than pointing to a deficit in reading one specific emotion (say, fear) or one specific channel (say, facial expressions), the data suggest systemic impairment across modalities — a generalized deficit in social cognition rather than a targeted one.

The most clinically significant result, however, was directional rather than scalar. Participants with a history of interpersonal aggression showed a negative to positive misclassification rate of 14% of cases — approximately twice the rate in healthy controls and 1.2 times the rate in non-aggressive psychosis patients (X²=53.26, p

The researchers interpret this pattern as a potential mechanism for violence escalation: if a person with psychosis is systematically reading another person’s fear, disgust, or anger as something neutral or positive, they lack the social feedback that would normally signal to stop, withdraw, or de-escalate. The confrontational partner’s distress is invisible to them — or worse, reads as encouragement.

What Predicts Emotion Recognition Ability?

Regression analysis across the full sample of psychosis patients identified four independent predictors identified of emotion recognition accuracy.

Fluid intelligence regression coefficient (β=0.364) — tested using the WAIS-IV Matrix Reasoning subtest, which measures novel problem-solving without relying on prior knowledge — emerged as the strongest single predictor, reinforcing the view that emotion recognition is a cognitively demanding perceptual task, not merely a reflexive social instinct. Semantic knowledge of emotion words, assessed via a vocabulary synonym task, was the second-strongest predictor (β=0.286), suggesting that the ability to conceptually differentiate emotion categories in language also shapes perceptual recognition in practice. Gender contributed significantly (β=0.176, women outperforming men), as did educational attainment (β=0.156).

A counterintuitive finding: a history of substance abuse was statistically associated with better ERAM performance once other factors were controlled for. The substance abuse discussion section explicitly cautions that this likely reflects a statistical artifact rather than a genuine protective effect of substance use, and should not be interpreted causally until replicated.

Critically, aggression history regression results showed that a history of interpersonal aggression independently predicted poorer emotion recognition even after intelligence, vocabulary, gender, education, and substance use history were all accounted for (β=0.177, p=0.027). Aggression history explained an additional 2.1% of the variance in recognition accuracy — a modest share of a complex picture, but a statistically meaningful and independent one. The implication is that something about having a history of aggression specifically — above and beyond the cognitive and educational deficits that often accompany psychosis — worsens the ability to read others’ emotional states.

Could This Test Become a Forensic Risk Assessment Tool?

The finding that aggression history predicts emotion recognition performance independently of general cognitive ability raises an underexplored clinical question: could a standardized emotion recognition test add measurable, reproducible value to forensic psychiatric risk assessment?

Current forensic risk tools, such as the HCR-20 (Historical-Clinical-Risk Management-20) and SAPROF (Structured Assessment of Protective Factors), operate through structured professional judgment tools — weighting dynamic clinical factors and historical variables based on trained clinician review. These tools have strong validity records but rely heavily on qualitative clinical judgment and historical documentation. A neuropsychological instrument that produces a quantitative score, can be administered consistently, and predicts aggression history independently of what a case file says would offer a meaningfully different kind of evidence.

“In summary, individuals with PSD and a history of interpersonal aggression exhibit more pronounced deficits in emotion recognition than those with PSD alone,” the authors wrote in the study conclusion and recommendations. “This underscores the potential value of incorporating emotion recognition assessment and training into clinical interventions to reduce aggression risk and improve social functioning in individuals with PSD.”

The study’s cross-sectional design means the current evidence cannot establish whether emotion recognition deficits causally precede violence, follow from the consequences of violent history (incarceration, social isolation, trauma), or reflect a shared neurological substrate underlying both. That causal question will require longitudinal follow-up studies — a recognized limitation the authors note explicitly.

Social Cognitive Remediation: An Underutilized Intervention

If impaired emotion recognition is a mechanism linking psychosis to violence rather than merely a correlate of it, then improving emotion recognition could theoretically reduce violence risk. A growing evidence base suggests this is not unrealistic.

Social cognitive remediation programs — structured therapies designed to improve the ability to read facial expressions, vocal tones, and body language — have shown promise in schizophrenia research. A 2024 study in Frontiers in Psychiatry found that a combined cognitive training and sleep intervention reduced aggression scores significantly in hospitalized schizophrenia patients after eight weeks, with treated patients scoring 3.15 on a standard aggression scale compared to 5.27 in a standard-care control group. Earlier research in European schizophrenia populations has also demonstrated that training programs targeting affect recognition can produce electrophysiological changes in recognition circuits consistent with improved perceptual processing, suggesting neuroplasticity in these circuits.

Despite this evidence, social cognitive remediation programs are not yet routinely integrated into forensic psychiatric care pathways. The Karolinska findings strengthen the case for changing that — and for developing training protocols specifically targeting the negative-to-positive misclassification error type, which this research suggests is the most clinically dangerous form of emotion recognition failure in this population.

Study Constraints and What Comes Next

The research team acknowledges study limitations acknowledged that future work will need to address.

The cross-sectional design is the primary one. Because all measurements were taken at a single point in time, no causal direction can be established. The data are consistent with emotion recognition deficits contributing to violence, with violent histories degrading emotion recognition, or with a shared biological substrate underlying both. Only prospective studies — tracking participants’ emotion recognition scores and subsequent behavior over time — can determine which.

The two psychosis groups differed substantially on group demographic differences table: more than three-quarters of the aggression group had a history of alcohol or substance abuse, compared with around a third of the non-aggression group. More than half of the aggression group had not progressed beyond primary-level education, versus 15% of the non-aggression group. The authors statistically controlled for these differences, but residual confounding cannot be excluded.

All testing was conducted during periods of clinical stability — a methodologically sound choice, but one that means the study cannot speak to how emotion recognition performs during acute psychotic episodes or agitation, which are precisely the states most associated with violent incidents.

The ERAM test’s stimuli were performed by white European actors based in Switzerland, limiting the generalizability of the results to other cultural and demographic groups. Future studies using more diverse actor pools would strengthen the evidence base significantly.

Finally, like virtually all current emotion recognition research, ERAM uses professionally performed rather than spontaneous emotional expressions — a limitation shared across the field and acknowledged by the authors as an open challenge for ecological validity.

A Decade of Stockholm Research Behind the Finding

This study represents one conclusion of a six-year multi-institution research program. The SPRätt project at Stockholm University, the Stockholm’s Project for Forensic Psychiatry, systematically compared neurocognitive, biological, physiological, and neuroimaging measures across matched forensic and non-forensic psychiatric populations between 2018 and 2024. Prior SPRätt publications have linked impaired emotion recognition in psychosis to processing speed deficits and lower educational attainment, and have examined the role of psychotically-driven aggression — violence motivated directly by delusions or hallucinations — in shaping social cognition outcomes. The new study is the first to isolate the negative-to-positive misclassification pattern as a specific feature distinguishing aggressive from non-aggressive psychosis patients.

Senior author Anette Johansson at Karolinska, a forensic psychiatrist and Docent at Karolinska Institutet who helped design and oversee SPRätt, has indicated that the next phase of the group’s work will explore whether machine learning applied to the full range of SPRätt’s cognitive and biological measures can sharpen risk prediction further. If emotion recognition scores are integrated into such models alongside neuroimaging, biomarker, and clinical data, the discriminative power could prove substantially greater than any single measure alone.

For now, the study stands as the largest and most methodologically comprehensive multimodal investigation to date of how psychosis and aggression history interact to shape the perception of other people’s emotions. Its central finding — that the specific error of reading negative emotions as positive is not merely an artifact of general cognitive decline but a pattern independently associated with violent history — gives forensic clinicians both a potential diagnostic signal and a concrete target for therapeutic intervention.

The Gavalova et al. 2026 paper, “Emotion recognition misclassification patterns in individuals with psychotic spectrum disorders and history of interpersonal aggression,” was authored by Gabriela Gavalova, Petri Laukka, Lennart Högman, Malin V. Källman, Marianne Kristiansson, Märta Wallinius, Håkan Fischer, and Anette G. M. Johansson.

Frequently Asked Questions

Why does misreading negative emotions as positive specifically increase violence risk — isn’t any emotion recognition error dangerous?

Not equally. Most emotion recognition errors in psychosis involve confusing one negative emotion for another — mistaking fear for anger, for instance — which still signals to the observer that something is wrong. The negative-to-positive error is uniquely hazardous because it removes the danger signal entirely: when a person reads another’s fear or anger as friendliness, they receive false social feedback that the interaction is going well. The normal cognitive “brake” — perceiving that the other person is frightened or hostile — never fires. The Karolinska team describes this as disabling the social circuit breaker that typically stops confrontational behavior before it escalates. No other error type produces this specific consequence.

Could a 15-minute emotion recognition test actually be added to forensic psychiatric risk assessments?

That is the clinical question this study raises most urgently. Current forensic risk tools like the HCR-20 and SAPROF are structured professional judgment instruments that weight clinical history and dynamic factors. They do not include standardized neurocognitive measures. The Karolinska data show that ERAM-style testing adds 2.1% unique predictive variance over and above all clinical and demographic factors combined — a modest but statistically independent contribution. Whether that addition is practically meaningful will depend on follow-on studies showing the score predicts future violence longitudinally, not just retrospective aggression history. The study authors call for that research explicitly, describing emotion recognition assessment as a candidate for integration into clinical intervention pathways.

Can people with psychosis improve their emotion recognition ability through training?

Emerging evidence suggests yes. Social cognitive remediation therapy — structured programs teaching patients to more accurately read facial expressions, vocal tone, and body language — has shown measurable effects in schizophrenia research. A 2024 Frontiers in Psychiatry study found that combined cognitive and sleep training reduced aggression scores significantly over eight weeks in hospitalized schizophrenia patients. Other research has found electrophysiological changes in recognition circuits in schizophrenia patients after affect recognition training, consistent with neural plasticity in relevant circuits. These programs are not yet standard in forensic psychiatric care, but the Karolinska findings strengthen the case for developing training protocols that specifically target the negative-to-positive misclassification error this research identifies as the clinically dangerous pattern.

Do these findings mean psychosis patients are inherently dangerous?

No — and the study is explicit on this point. Psychotic disorders carry a statistically elevated violence risk compared to the general population, but the absolute risk remains low, and the vast majority of people living with schizophrenia or related conditions will never commit a violent act. The study’s purpose is precisely the opposite of stigmatization: by identifying the specific neurocognitive mechanism linking psychosis to violence risk in a small subset of patients, it creates a path toward better-targeted clinical intervention rather than blanket risk classification. The finding that aggressive and non-aggressive psychosis patients differ on a measurable cognitive dimension — not just on diagnosis or symptom severity — supports a precision-medicine approach to forensic psychiatric care.