Artificial intelligence for the diagnosis and treatment response prediction of obsessive-compulsive disorder: A narrative review

Abstract

Obsessive-compulsive disorder (OCD) exists as a persistent psychiatric condition which produces different symptoms that lead to severe personal distress and complete social and work disability and major expenses for communities. The field of neurobiology has made progress but doctors still face two main clinical obstacles which include delayed diagnoses and inconsistent treatment outcomes and numerous cases of medication failure. The field of psychiatry now benefits from two separate developments which include artificial intelligence (AI) advancements and psychiatric treatment method improvements. This study is a narrative review that synthesizes existing evidence on the application of AI in OCD, with a focus on diagnostic models, treatment response prediction, and clinical decision support systems. The research combines existing data about AI usage in OCD treatment through its evaluation of diagnostic methods and treatment prediction models and clinical decision systems. The research team performed a thorough evaluation of studies which used machine learning and deep learning models to analyze neuroimaging data and electrophysiological signals and clinical scales and digital phenotyping information and multiple data sources. The research evaluated two new roles of large language models together with explainable AI methods which assess their potential for medical interpretation and their ethical implications and their ability to translate into practice. The research findings show that AI-based models demonstrate successful results in identifying OCD patients from healthy participants and in forecasting treatment outcomes for pharmacological and psychotherapeutic and neuromodulation interventions. Research shows that models which use multiple data sources produce better results than systems which depend on a single data source because OCD exists as a complex combination of neurological and behavioral elements. The reported study results show significant differences between research investigations because the studies used different data sources and testing methods and validation approaches. The current limitations in explainability together with insufficient external validation and small available data samples prevent these models from being used in clinical practice. AI provides substantial potential to enhance precision psychiatric care for OCD patients through its ability to detect conditions earlier and create personalized treatment plans and provide continuous clinical guidance. The process of translating this technology into medical practice needs research studies which must be conducted in multiple centers throughout different time periods while using AI systems that provide explanations and follow human-friendly design principles. The deployment of AI tools for obsessive-compulsive disorder management requires solutions to address ethical issues and regulatory requirements and interpretability problems for safe and dependable and enduring implementation.

Key Words: Obsessive-compulsive disorder; Artificial intelligence; Machine learning; Deep learning; Clinical decision support systems; Precision psychiatry; Neuroimaging; Treatment outcome; Explainable artificial intelligence

Core Tip: This review highlights that artificial intelligence (AI) enables earlier and more accurate obsessive-compulsive disorder identification by integrating multimodal biomarkers (magnetic resonance imaging, electroencephalography, clinical scales, and digital phenotyping), surpassing unimodal models. Its key innovation is the joint clinical role of explainable AI for transparent neurobiological interpretation and large language models for personalized, scalable clinical decision support. Limited explainability, small samples, and weak external validation remain the main barriers to clinical translation. Future obsessive-compulsive disorder care requires multicenter, longitudinal, and clinician-aligned explainable AI systems to ensure ethical, regulatory, and trustworthy implementation.

INTRODUCTION

Obsessive-compulsive disorder (OCD) exists as a long-lasting psychiatric condition which presents differently to each person through intrusive thoughts and repetitive behaviors that create major life difficulties and decrease overall life satisfaction. The clinical definition of OCD includes two main components which are obsessions that produce intrusive distressing thoughts and images and compulsions which are repetitive mental or physical actions that people perform because of their obsessions[1]. Research findings show that OCD symptoms emerge from neurobiological brain circuit problems which affect the Cortico-Striato-Thalamo-Cortical (CSTC) system and from imbalanced neurotransmitters and multiple symptom patterns. Research shows that the disorder’s core symptoms result from dysregulation of the brain circuitry which includes the orbitofrontal cortex (OFC) and anterior cingulate cortex and striatum[2-4].

The diagnosis and treatment of OCD continue to present difficulties to medical professionals despite the progress made in neuroimaging and psychopharmacology. The clinical process becomes difficult because patients show different symptoms and their diseases progress at different rates and their bodies react differently to treatments. The combination of serotonergic deficits with dopaminergic and glutamatergic hyperactivity in CSTC loops creates additional complexity in OCD pathophysiology because of neurotransmitter dysregulation. The complexity of OCD pathophysiology requires individualized assessment and intervention models which should integrate all available information[5].

The current situation with OCD treatment faces multiple obstacles which stem from these challenges. Machine learning (ML) and deep learning (DL) techniques enable researchers to analyze complex high-dimensional and multimodal data which helps them detect faint neurobiological and behavioral patterns that standard statistical methods cannot detect. The research of OCD has started using these methods to develop new diagnostic methods which help scientists identify treatment responses and create symptom-based categories for better psychiatric care[6,7].

Artificial intelligence (AI) represents a complete scientific discipline which uses various computational methods to create systems which perform tasks that humans can do through learning and reasoning and pattern detection. The system contains ML as its AI component which uses data algorithms to discover statistical patterns without needing programmed rules. DL operates as a ML subfield which uses multiple neural network layers to analyze complex nonlinear data patterns. The application of support vector machines and least absolute shrinkage and selection operator regression in psychiatric research enables the analysis of complex neuroimaging and electrophysiological and clinical data through successful feature selection and classification methods. The analysis of complex neurobiological and behavioral patterns in OCD becomes possible through DL approaches which include convolutional and graph-based neural networks that automatically discover hierarchical features from multiple data sources.

The main benefit of AI-based methods stems from their ability to process information from multiple data types. Neuroimaging modalities provide brain network information through their ability to show both structural and functional brain network details[8]. The brain’s dynamic neural activities become measurable through electroencephalography (EEG) recordings which provide electrophysiological data. Digital phenotyping allows researchers to monitor people’s actual behaviors and their symptoms which change throughout the day. The combination of these data streams which work together has proven to boost model accuracy while making the models more responsive to personal differences which are essential for understanding OCD’s diverse nature[7,9].

The process of turning AI technology into clinical applications for OCD treatment encounters multiple major obstacles which include methodological challenges and ethical concerns and practical implementation difficulties. The generalization of results becomes limited because datasets show different characteristics and researchers work with small numbers of participants while using different validation methods. The unexplainable operation of advanced AI systems which achieve top results creates difficulties for medical staff to understand their decisions while building trust and upholding ethical standards in psychiatric treatment. The development of explainable AI (XAI) has become essential because it provides model decision transparency which produces clinical outcomes and solves problems with bias and safety and regulatory needs[6,10].

The current review evaluates AI applications for OCD through a critical assessment of existing research which includes neurobiological mechanisms and diagnostic and prognostic systems and individualized treatment approaches and explainable systems. The review combines clinical neuroscience findings with computational psychiatry data to establish the present status of OCD treatment with AI while it identifies essential research challenges and develops necessary paths for AI implementation in OCD treatment.

Consistent with the aims of a narrative review, this article does not attempt a quantitative synthesis or formal risk-of-bias assessment. The review conducts a conceptually based critical evaluation of vital research evidence which enables developers to create AI methods for OCD treatment and identifies operational barriers for their implementation.

METHODOLOGY

The research used narrative review methods to combine existing evidence about AI treatment for OCD. The research required a qualitative method because the field contains diverse methods and fast-changing technology and multiple types of data that cannot be analyzed through systematic or meta-analytic methods.

The research team conducted specific database searches to identify relevant studies which appeared in PubMed/MEDLINE and Web of Science and Scopus and IEEE Xplore. The research investigation concentrated on scientific articles which appeared between 2010 and 2025 because this time frame saw ML and DL and XAI methods start their use in psychiatric studies. The research used search queries to link OCD terms with AI approaches which included ML and DL and neuroimaging and EEG and treatment response prediction and clinical decision support systems and digital phenotyping and large language models (LLMs) and XAI. The research team performed manual screening of reference lists which contained essential articles and current review publications to discover additional relevant studies.

The research team used clinical and translational AI applications in OCD to select studies which focused on diagnostic classification and treatment response prediction and neuromodulation strategies and digital therapeutics and crisis monitoring and explainability. The review process evaluated original research articles from peer-reviewed journals together with systematic reviews and studies which used AI methods to produce clinical results and various data types. The research team excluded studies which presented conference abstracts without full-text access and single-case reports and studies that did not provide adequate methodological information.

The research used narrative review as its method because it did not perform any quantitative synthesis or meta-analysis. The research team applied clinical psychiatry and computational science approaches to their findings through thematic synthesis. The research focused on three main areas which included data types and processing methods and validation techniques and performance results and the restrictions that affect the methods. The research method enabled experts to evaluate existing studies while they identified three major evidence-related problems which stem from insufficient study participants and insufficient model testing outside the original data and difficulties in understanding model operations and obstacles to using the results in medical practice.

EPIDEMIOLOGY

Epidemiological findings in OCD have varied across studies due to differences in diagnostic criteria, classification approaches, and the clinical heterogeneity of sampled populations[11].

The National Comorbidity Survey documented that 2073 participants showed obsessions or compulsions during their lifetime at a rate of 25%. The research study found that OCD occurred in 1.2% of participants during the 12-month period based on Diagnostic and Statistical Manual of Mental Disorders, fourth Edition (DSM-4) diagnostic standards[11,12]. Research studies initially indicated OCD affected people at a minimal rate but subsequent studies revealed that OCD affects 2%-3% of people throughout their lives while showing different occurrence patterns across different areas[1,13]. The majority of OCD cases start during young adulthood when people reach their early twenties[1]. The first symptoms of the condition appear before boys reach age 10 but girls usually start showing symptoms during their teenage years[14].

The likelihood of being affected by OCD has been reported to be approximately 1.6 times higher in women than in men[15]. Consistent with these findings, pooled analyses from the World Health Organization’s international surveys have estimated the lifetime prevalence of OCD to be 2%-3%[16]. Subthreshold OCD symptoms, however, have been reported at substantially higher rates[16,17]. In a large prospective study conducted across ten countries with 26136 participants, the 1-year prevalence of OCD was reported as 3%, while lifetime prevalence reached 4.1%; moreover, age at onset was identified as early adulthood in approximately 80% of cases[17]. Across the literature, similar patterns of sex differences have been observed, with earlier onset during childhood in males and onset during adolescence or young adulthood in females. In addition, symptom exacerbation during pregnancy and the postpartum period has been reported among women, who also show higher rates of comorbid anxiety and depressive disorders[18].

The social structure of society faces a major economic challenge because OCD creates two types of problems which affect both medical treatment of the condition and its effects on human beings (Figure 1)[19,20]. The therapy expenses for patients have been estimated to fall between United State dollar 50000 and 150000 per person[21]. Research studies which lack direct monetary assessments demonstrate that OCD creates substantial social costs because it reduces work capacity and requires patients to spend time on their symptoms and leads to increased costs for their treatment[20,22].

Figure 1 Economic burden of the disorder. AI: Artificial intelligence; CBT: Cognitive-behavioral therapy; CNN: Convolutional neural network; EEG: Electroencephalography; ERP: Exposure with response prevention; fMRI: Functional magnetic resonance imaging; GCN: Graph convolutional network; LIME: Local interpretable model-agnostic explanations; OCD: Obsessive-compulsive disorder; SHAP: SHapley Additive exPlanations; sMRI: Structural magnetic resonance imaging; SSRI: Selective serotonin reuptake inhibitor; SVM: Support vector machine; TMS: Transcranial magnetic stimulation; XAI: Explainable artificial intelligence.

The consistent occurrence of this condition across different cultures requires researchers to conduct extended studies which combine complete data collection from multiple research sites.

ETIOLOGY

The etiology of OCD has not been fully elucidated; however, multiple theoretical models have been proposed to explain its development[23]. The models of PTSD include neurobiological theories which show CSTC circuit dysfunction[24,25] and behavioral learning models which study fear acquisition and extinction processes[26] and cognitive models which focus on abnormal beliefs and incorrect interpretation patterns[27,28]. The existing frameworks of OCD treatment include three main approaches which are the cognitive-behavioral model and the neurobiological model and the psychodynamic model. The three main cognitive deficits which affect OCD patients include their inability to change their thoughts and their failure to control their responses and their problems with achieving their desired outcomes[23].

The neurobiological evidence shows that parallel CSTC circuits function as primary factors which lead to the development of OCD[24,25]. These circuits play essential roles in reward and motivation systems and executive functions and motor control and response inhibition and habit-based behaviors[29]. Research using structural neuroimaging techniques has shown that CSTC components show different patterns of change which include decreased volume of the OFC[30]. Research using AI analytical methods has produced identical results which show that the dorsolateral prefrontal cortex and anterior cingulate cortex and OFC and amygdala experience structural changes[6,31]. Research conducted by AI models demonstrated that these neural substrates play a crucial role in OCD pathophysiology because they achieved 86.4% accuracy in predicting treatment responses between pharmacological and psychotherapeutic approaches[32].

The established efficacy of selective serotonin reuptake inhibitors (SSRIs) in OCD treatment has confirmed that the serotonergic system plays a key role in understanding OCD origins[33]. Research on treatment methods shows that SSRIs work in OCD by changing the activity of the OFC[34] according to studies which AI-based research has also verified through its ability to detect SSRI-induced changes in OFC function[6,35]. Research based on clinical observations shows that antipsychotic drugs which target dopamine pathways help treat OCD symptoms which indicates dopamine system involvement in OCD development[36]. The excitatory neurotransmitter glutamate functions as the main excitatory signal which runs through CSTC pathways and neuroimaging studies show that glutamatergic activity rises in particular brain areas of OCD patients[37]. Research indicates that OCD develops from neurot-ransmitter system imbalances together with changes in brain structure that affect particular brain areas[23] while AI-based research methods continue to confirm these findings[6,32,35].

The learning model provides an alternative method to explain how neutral stimuli that accompany unpleasant events result in fear reactions and obsessive behaviors[26].

In contrast, cognitive models emphasize the role of maladaptive beliefs, proposing that compulsive behaviors are repeatedly enacted to reduce distressing thoughts or anxiety and to prevent perceived negative outcomes[27,28]. AI-based studies designed to operationalize these theoretical models have demonstrated the therapeutic relevance of cognitive-behavioral interventions, including cognitive restructuring strategies and exposure and response prevention, in the treatment of OCD[38,39]. Beyond these models, additional factors such as psychosocial stressors, obstetric complications, family history, and stressful life events have also been shown to contribute to the etiology of OCD[40,41].

Clinical features, symptoms, and diagnostic criteria

According to the DSM-5, OCD is classified under the category of Obsessive–Compulsive and Related Disorders (Table 1)[42].

Table 1 Obsessive-compulsive and related disorders.

Number	Disorder
1	Obsessive-compulsive disorder
2	Body dysmorphic disorder
3	Hoarding disorder
4	Trichotillomania (hair-pulling disorder)
5	Excoriation (skin-picking) disorder
6	Hypochondriasis (ICD-11)
7	Tourette syndrome (ICD-11)
8	Olfactory reference syndrome (ICD-11)

ICD: International Classification of Diseases.

Obsessions are defined as recurrent, persistent, and intrusive thoughts, images, urges, or impulses that arise involuntarily and are experienced as distressing, typically in association with anxiety. Compulsions refer to repetitive behaviors or mental acts that an individual feels driven to perform in response to an obsession[17]. Common obsessional subtypes include contamination, aggression, symmetry, doubt, and religious or sexual obsessions, whereas frequently observed compulsions encompass cleaning, checking, repeating, ordering, and hoarding behaviors[43]. A study reported in the literature demonstrated that AI-based systems were able to distinguish individuals with OCD from healthy controls based on symptom profiles with an accuracy of up to 98%. In terms of discriminative power, contamination and cleaning, symmetry, aggression, and religious and sexual obsessions were identified as the most informative symptom dimensions[44].

With the DSM-5 diagnostic and classification system, specifiers related to the level of insight were introduced for OCD, distinguishing individuals who recognize that their beliefs are definitely or probably not true (good or fair insight), those who believe that their obsessive beliefs are probably true (poor insight), and those who are completely convinced that their beliefs are true, reflecting absent insight or delusional beliefs[17].

Treatment

Serotonergic agents, including clomipramine, fluvoxamine, fluoxetine, and paroxetine, have long been extensively studied and are widely recommended pharmacological treatments for OCD[45]. In the literature, SSRIs are consistently identified as the first-line treatment protocol for OCD[45,46]. Compared with the treatment of depressive disorders, SSRIs are typically administered at substantially higher doses and for longer durations in OCD[47]. Among tricyclic antidepressants, clomipramine has been shown to be effective as a second-line treatment and, in some studies, to be more efficacious than SSRIs[48].

Pharmacological augmentation strategies are frequently employed, most commonly involving antipsychotic agents or the addition of antidepressants from other pharmacological classes. During augmentation therapy, careful attention must be paid to medication adherence, tolerability, and the potential for adverse effects[49-51]. In addition to pharmacological approaches, psychotherapeutic interventions, either as adjuncts to medication or as standalone treatments, constitute a central component of OCD management[45].

Electroconvulsive therapy

Electroconvulsive therapy (ECT) is a treatment modality that induces a controlled epileptic seizure through intracranial electrical stimulation. ECT is used in the treatment of a wide range of psychiatric disorders, and evidence suggests that it may be effective in patients with OCD who do not benefit from conventional treatment algorithms[52-54]. ECT is used with general anesthesia. Although sufficient muscle relaxation is necessary during ECT, forceful jaw clenching is still inevitable with this intervention because of the direct stimulation of the masticatory muscles, particularly the temporalis and masseter muscles, by electrical current. Thus, a bite block should be carefully placed before the application of the electrical stimulus to protect the patient’s teeth and minimize the risk of lacerating the tongue. The overall magnitude of skeletal and dental fracture risk depends on patient factors and on how well the ECT procedure is performed. Preexisting bone and dental disease increase the risk; good seizure modification, proper use of bite blocks, and effective jaw immobilization during ECT reduce the risk. Careful assessment of preexisting risk and good ECT practice can minimize the risk of skeletal and dental complications during ECT[55,56].

Transcranial magnetic stimulation

Transcranial magnetic stimulation (TMS) is a noninvasive brain stimulation technique with well-established efficacy in major depressive disorder[57]. In patients with OCD, treatment outcomes have been shown to vary depending on the specific brain region targeted for stimulation. Overall, findings reported in the literature indicate that TMS is generally associated with a reduction in symptom severity in individuals with OCD[58,59].

Personalized therapeutic approaches

Since Meyer[60] first described exposure and response prevention in 1966, numerous adaptations and refinements of cognitive-behavioral therapy (CBT) have been developed and implemented in the treatment of OCD[61]. Exposure with response prevention (ERP) is based on the systematic and hierarchical exposure of individuals to their obsessions while preventing the execution of compulsive behaviors. Through this process, patients are expected to experience a manageable level of distress and to develop the capacity to cope with anxiety without resorting to compulsions[61]. Cognitive interventions targeting dysfunctional beliefs are also integral components of CBT for OCD[62]. In addition, acceptance and commitment therapy, which is classified as a third-wave therapeutic model and is widely used in OCD, has been reported to be highly effective in clinical studies[63].

Course, outcome, and prognosis

A review of the literature indicates that long-term remission can be achieved when an accurate diagnosis is established, an appropriate treatment model is implemented, and adequate treatment adherence is maintained[64,65]. In a study conducted among individuals with OCD characterized by severe obsessions, symptom improvement was observed in approximately half of the participants over a five-year follow-up period. In this cohort, treatment adherence was found to deteriorate most frequently within the first three years, while the remaining half of the participants continued to exhibit persistent symptoms over time[66].

Several factors have been identified as key determinants of prognosis in OCD, including age at onset, sex, level of insight, adherence to treatment, early engagement in treatment, and the presence of comorbid psychiatric disorders[66]. Younger patients face three main obstacles which include work-related challenges and expensive medical care and their inability to perform daily activities outside their home. The disorder requires immediate identification because individualized evidence-based treatment methods need to be started right away to achieve better results and decrease the total weight of the condition on people and communities[20,22].

Overview of AI applications in OCD

In recent years, AI applications in OCD have emerged as a comprehensive clinical support domain encompassing diagnosis, treatment planning, prediction of treatment response, and long-term monitoring processes. The combination of ML approaches with DL and natural language processing and augmented reality and generative AI allows researchers to create detailed models of OCD clinical diversity while enabling them to study biological and behavioral markers together and building individualized clinical decision-making tools. Research studies demonstrate that AI diagnostic systems which analyze neuroimaging data together with clinical rating scales and behavioral measurements achieve diagnostic accuracy between 21% and 100% for mental disorders; the broad range of results stems from differences in dataset dimensions and annotation precision and algorithm intricacy[67].

ML algorithms have proven their ability to forecast OCD treatment outcomes which enables healthcare providers to create personalized medication plans and determine the best drug amounts and duration of treatment and to detect patients who will not respond to treatment[6]. The therapeutic use of AI-based neuromodulation methods with closed-loop deep brain stimulation technology allows doctors to adjust stimulation settings through biological feedback which produces effective results for patients with treatment-resistant OCD[6]. AI-based virtual therapy platforms and neurofeedback applications provide better access to care through their digital platforms which help patients maintain their treatment plan by organizing ERP-based interventions in virtual spaces. The development of augmented reality applications enables healthcare professionals to establish protected environments which help patients with contamination-related OCD experience controlled exposure to their fears while therapists monitor their emotional responses and reward their learning progress[68].

The assessment of AI-based intervention clinical effectiveness requires more than technical performance indicators which include accuracy and sensitivity and specificity. The assessment framework needs multiple layers which should evaluate user experience together with crisis management and clinical reliability and transparency and accountability. The FAITA-Mental Health framework enables researchers to conduct thorough assessments of AI mental health systems through its three evaluation criteria which include scientific validity and user-centeredness and ethical compliance. The framework shows that generative AI applications for OCD provide good user interaction and easy access but it reveals major requirements to enhance crisis response systems and medical practice integration and follow regulatory standards[69].

Digital phenotyping through smartphone usage patterns, speech analysis, and real-time symptom monitoring may help identify sudden symptom exacerbations, suicidal ideation, or functional deterioration. Chatbot-based systems and virtual mental health platforms can provide immediate psychoeducation, emotional support, and guidance toward professional help during acute distress[70].

However, current AI-based crisis intervention tools are still limited by insufficient validation in OCD-specific populations, ethical concerns regarding data privacy, and the risk of misclassification of high-risk individuals. False reassurance or delayed emergency referral may pose serious clinical risks. Therefore, AI systems should be viewed as adjunctive tools rather than replacements for clinical judgment, and integration with supervised clinical pathways is essential[71].

Future research should focus on developing disorder-specific crisis detection algorithms, improving sensitivity to suicidal risk in OCD, and evaluating the real-world effectiveness of AI-assisted crisis management strategies. The implementation of AI applications for OCD faces three significant ethical obstacles which stem from using sensitive mental health information and require algorithmic decision transparency and result from unbalanced dataset distribution. The implementation of XAI methods together with enhanced privacy and data protection standards and human involvement in AI-based clinical choices will create conditions for trustworthy AI integration in future OCD treatment[6,72].

Digital phenotyping which analyzes smartphone usage patterns and speech data and tracks symptoms in real-time enables the detection of unexpected symptom worsening and suicidal thoughts and decreased functional ability. Users can find psychoeducational content and emotional support through chatbot systems and virtual mental health platforms which will guide them to professional assistance when their distress reaches severe levels. The current AI-based crisis intervention tools face three major restrictions because they lack proper validation for OCD patients and they violate privacy rules and they fail to identify people who need immediate help. The practice of false reassurance or delayed emergency referral creates dangerous medical situations for patients. AI systems should operate as medical decision support tools which assist doctors without taking their place while needing clinical pathway supervision for correct deployment. Research efforts should focus on developing crisis detection systems which focus on particular mental health conditions and improving their capacity to detect suicidal risks in OCD patients and conducting real-world clinical tests of AI-based crisis management systems. The safe deployment of AI in OCD crisis management systems depends on XAI methods and improved privacy protection and continuous human monitoring.

The five stages of AI-based OCD management exist as interconnected elements which operate within a precision psychiatry framework according to Figure 2. The system starts with ML and DL models which process multimodal neurobiological data to achieve both diagnostic and classification results. The stage 2 system uses combined clinical and neuroimaging assessment to determine how patients will react to pharmacological treatments and psychotherapeutic approaches and neuromodulation therapies. Stage 3 employs LLMs to create individualized treatment plans which also offer clinical decision support to users. Stage 4 implements continuous monitoring and crisis intervention via digital phenotyping and wearable technologies. The system achieves explainability and clinical trust at stage 5 through XAI methods which include SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) and attention mechanisms. The lower section demonstrates how these applications work together as a single system because diagnostic models use their output to choose treatments while multimodal data integration improves prediction results and explainability frameworks support ethical deployment throughout the complete healthcare process.

Figure 2  Conceptual framework of artificial intelligence applications in the obsessive-compulsive disorder care pathway.

ML and DL-based classification models in the diagnosis of OCD

Research on ML and DL-based classification models for the diagnosis of OCD has gained substantial momentum in recent years due to the disorder’s global prevalence, chronic course, and pronounced negative impact on quality of life[73,74]. Affecting approximately 1%-3% of the population worldwide, OCD is characterized by intrusive thoughts and repetitive behaviors and has been consistently associated with functional and structural abnormalities across distributed brain networks, particularly within CSTC circuits[75,76].

Neuroimaging technology progress during the last ten years together with computational method development enabled scientists to discover neurobiological biomarkers which are specific to OCD. Research conducted during the early period focused on using structural magnetic resonance imaging (MRI) and functional MRI data[74,77]. The current research combines ML with DL to create new predictive models which analyze each person separately for better diagnostic results and customized risk assessment[6,78]. The trend presents important medical and public health implications because doctors face multiple difficulties when they need to identify OCD through early and precise methods[79,80].

The development of dependable ML/DL models which can classify OCD continues to be an unsolved problem. The literature shows that reported classification accuracy results span from 57% to 98% but researchers have not established reliable results for different study groups and clinical sites and data types[73,74,81]. The current state of research lacks methods to unite different types of data and to establish common procedures for feature selection and model verification using extensive datasets from multiple research sites[73,82,83]. The medical community continues to discuss which neuroimaging technique provides the best results and which ML and DL system works best for OCD identification. Research studies have shown that EEG-based methods provide better temporal resolution and portability according to Farhad et al[84] supports this view but other researchers[85] believe MRI-derived structural and functional characteristics deliver better spatial precision according to Bu et al[75] and Huang et al[86]. Research investigations now show that genetic data together with multimodal approaches provide superior methods to understand the different clinical and neurobiological presentations of OCD[87,88]. The medical field continues to debate about how well DL models can be understood by doctors which creates obstacles for using these models in clinical practice for decision support systems and for implementing early intervention and individualized treatment approaches[6,10,81].

The framework uses ML/DL to diagnose OCD by identifying which neurobiological data points differentiate people with OCD from those without the condition[10,78]. The method requires three essential elements which consist of brain functional connectivity measurements and neuroimaging-based biomarkers and support vector machines and convolutional neural networks and graph convolutional networks for classification purposes[89-91]. The system connects neurobiological changes to feature selection processes which ML/DL algorithms use to generate diagnostic predictions thus creating a direct path between computational techniques and medical results[74,82].

The main objective of this section involves analyzing current research about ML and DL classification models for OCD diagnosis to identify major weaknesses which stem from data inconsistencies and model performance on different datasets and model ability to understand results. The research synthesis from this section will direct scientists toward new investigation paths[73,83]. The evaluation of research data from neuroimaging and EEG and genetic and multiple study types will help scientists understand both new diagnostic methods and ongoing challenges which affect OCD diagnosis through AI-based systems[6,82].

Prediction of treatment response using AI

Predicting treatment response in OCD has emerged as a critical area of research in clinical psychiatry due to the disorder’s marked clinical heterogeneity, chronic course, and the high proportion of patients who fail to respond to first-line treatments. The literature indicates that approximately 40%-60% of individuals with OCD do not achieve an adequate response to standard interventions such as CBT/ERP and SSRIs[6,92]. The large number of nonresponding patients requires researchers to develop data-based methods which predict treatment results in advance thus AI models emerge as a suitable replacement for current clinical trial methods.

Research findings show that ML and DL algorithms use clinical data and neuropsychological information and neuroimaging results and physiological measurements to achieve accurate treatment prediction[7,78]. Research studies which unite structural and functional MRI data with EEG band power measurements and symptom severity ratings and patient demographics have achieved area under the receiver operating characteristic curve values between 0.70 and 0.90[6,93]. The research demonstrates that support vector regression models which analyze cortical structure data produce significant results for CBT treatment prediction. The combination of EEG beta-band activity with multimodal EEG-MRI integration produces better results than using single-modality approaches for deep TMS response prediction[93-96].

The main benefit of AI-based methods stems from their ability to handle OCD complexity by uniting different types of data into a single model. The combination of clinical symptom profiles with executive function measures and neuroimaging biomarkers and wearable sensor-derived physiological data allows better prediction of treatment outcomes for pharmacological and psychotherapeutic interventions[7,97]. Executive function impairments serve as predictors for pharmacological treatment response according to this framework while symptom network analysis models demonstrate how they show treatment-resistant patterns and comorbid conditions[98,99].

AI applications for treatment response prediction now extend past pre-dictive models to include digital interventions which help patients follow their treatment plans better and achieve better results. Real-time symptom tracking through AI-based digital CBT platforms and smartphone applications and wearable devices helps users monitor their symptoms continuously while detecting early signs of relapse and delivering immediate interventions[97,100]. The combination of peer-supported and AI-guided teletherapy applications leads to better treatment participation and patient compliance which produces substantial symptom reduction[100,101]. The long-term success of digital applications and user retention sustainability has not received enough evidence-based validation.

The current research faces multiple obstacles because of its methodological and clinical restrictions. Research studies currently use three main limitations which include working with small groups of participants and studying past events and studying patients who have similar health conditions without any additional medical conditions. The models developed from these studies fail to generalize to new situations because of these research design choices[78,96]. The process of clinical application faces three main obstacles which include performance differences between research centers and missing standardized data collection methods and insufficient studies that validate results over time[102,103].

The ethical and operational aspects of data privacy and algorithmic transparency and bias risk and clinician-patient relationship maintenance create significant problems[6,69]. The implementation of continuous monitoring systems through wearable technology and high-dimensional neurobiological data requires organizations to establish strong data governance systems and specific ethical guidelines. The current discussion shows that AI systems should assist medical professionals during clinical decision support operations instead of making independent decisions[104,105].

AI-based models show great potential for OCD treatment response prediction because they allow doctors to create individualized treatment plans based on biomarkers through analysis of multiple data types. The implementation of safe and effective clinical practice requires researchers to conduct multiple large-scale studies which span different centers and follow patients throughout time while they create official translation methods and enhance both ethical and regulatory systems. The expected core application of precision psychiatry for OCD treatment response prediction will depend on AI systems operating under these specific conditions.

LLMs and clinical decision support systems in personalized therapy planning

The integration of LLMs into clinical decision support systems for OCD offers the potential for a paradigmatic shift in personalized therapy planning. By virtue of their ability to semantically and contextually analyze large volumes of unstructured text data such as clinical interview notes, patient self-reports, digital diaries, and online interaction records LLMs provide significant advantages over traditional assessment approaches in capturing the pronounced clinical heterogeneity of OCD. In particular, LLM-based systems are increasingly being employed to identify symptom triggers, enhance diagnostic accuracy, and structure individualized therapeutic interventions.

The analysis of free-text data from OCD individuals by LLMs enables the identification of cognitive elements and emotional responses and environmental stimuli which lead to obsessive thoughts and compulsive behaviors. The system uses semantic similarity and intensity and contextual patterns to perform trigger classification. The method allows researchers to identify specific symptoms in each person which becomes essential for creating individualized treatment plans. Research has shown that LLM-based analyses help scientists detect multiple trigger types which occur in various ecological settings thus enabling them to create specific psychotherapeutic treatment plans[106]. The clinical decision support systems use LLMs as additional cognitive tools which help therapists create more specific treatment plans for their patients.

Research studies have shown that LLMs achieve high diagnostic accuracy when they evaluate OCD cases through clinical vignettes and symptom narratives to the point where their results match or exceed those of mental health experts. The research results indicate that the proposed method has the ability to decrease both diagnostic delay times and incorrect diagnosis rates which frequently occur in OCD patients[107]. The diagnostic systems of LLMs function to enhance medical assessments through their ability to perform detailed symptom severity assessments and comorbidity pattern identification and insight level evaluation. The combination of methods enables healthcare providers to create individualized treatment strategies through prompt identification of conditions at their beginning stages.

The development of LLMs for use in conversational AI platforms which deliver personalized therapeutic interventions continues to be an actively developing field. The delivery of evidence-based psychotherapeutic approaches through CBT and MI-based systems enables LLMs to create flexible dialogue patterns which produce individualized assistance for each user. The research shows that LLMs produce therapeutic content which users can evaluate to create personalized interventions[108]. The clinical value of LLM outputs becomes more appropriate when these outputs follow specific therapeutic frameworks which include CBT protocols[107]. The digital therapeutic interventions become more scalable through this alignment which maintains their clinical effectiveness.

The implementation of LLMs into clinical decision support systems creates major ethical problems together with operational difficulties. The sensitive nature of mental health information requires organizations to establish robust systems which protect both patient privacy and maintain complete data security. The clinical risks from LLMs become more severe because these systems produce fake information which requires healthcare personnel to maintain complete transparency while they need to monitor system performance at all times[109]. AI system ethical standards development requires AI developers to collaborate with mental health professionals who will develop guidelines and create clinical responsibility frameworks[110].

LLMs have proven themselves as essential instruments which help healthcare providers create personalized treatment plans for OCD patients while delivering important benefits for symptom evaluation and diagnostic processes and therapeutic program development. The essential principle requires these technologies to function as tools which assist human therapists during their decision-making activities while enhancing their ability to build strong therapeutic relationships. The implementation of LLM-based clinical decision support systems for mental health care requires innovative solutions together with ethical responsibilities to achieve stable and enduring results[109,110].

The diagnostic accuracy and clinical equivalence of human specialists to LLMs and digital therapeutic platforms needs complete evaluation because their performance was tested through experimental studies conducted in artificial clinical environments. Research findings from most studies emerged from studies conducted through retrospective methods and controlled vignette experiments and non-clinical testing environments instead of actual clinical practice settings. The current evidence shows these systems function best as experimental tools which provide additional support but should not be used as independent diagnostic or therapeutic methods. The distinction between proof-of-concept studies and simulated clinical evaluations and validated real-world applications needs clear definition because it prevents researchers from exaggerating the current clinical readiness of large language model-based interventions.

XAI in OCD

XAI in OCD has emerged as a fundamental paradigm for enhancing the trustworthiness and clinical acceptability of AI applications in psychiatric disorders characterized by marked heterogeneity, chronicity, and high levels of uncertainty in clinical decision-making processes. The worldwide occurrence of OCD affects 2%-3% of people yet patients commonly wait multiple years before starting proper treatment so there exists an escalating requirement for clinical decision support systems which provide early and specific and individualized assistance[71,73]. Research studies from the last few years have shown that ML and DL models which combine EEG data with neuroimaging information and clinical evaluation results and digital monitoring data and wearable device readings have achieved substantial diagnostic precision for OCD and they can measure symptom intensity and forecast treatment outcomes[6,7,97]. The majority of these models operate as black boxes because their decision mechanisms remain uninterpretable for clinicians which scientists identify as the main obstacle to their adoption in standard medical care[111,112].

The current stage of XAI development focuses on building clinical trust through methods which show users how models arrive at their results and what factors influence their decisions and prediction uncertainty. The clinical value of XAI emerges because it enables healthcare professionals to convert model results into practical information which helps them make diagnostic decisions and develop treatment strategies. The SHAP and LIME methods for feature attribution help researchers identify which neuroimaging markers and symptom dimensions and electrophysiological features affect the risk classification and treatment response predictions for patients. Healthcare providers in clinical environments check model suggestions against their scientific knowledge of brain functions and their personal observations of patient symptoms through explanatory processes.

Studies have shown that attention-based and hybrid XAI models can detect important clinical patterns which reveal how patients are affected by changes in their CSTC circuits and their executive function abilities. The process of model uncertainty disclosure along with decision pathway presentation has become essential for obtaining regulatory approval and ethical deployment of systems in psychiatric care because it ensures both accountability and patient safety. The examples show how explainability outputs serve dual purposes as they provide technical diagnostic information and clinical tools which help establish trust and meet regulatory needs and allow for safe AI deployment in mental health practice.

The current research shows that XAI studies about OCD lack sufficient numbers and post hoc explainability methods including SHAP and LIME have not received proper standardized implementation[73,113]. Research studies have proven that attention mechanisms work with Shapley-based explanations to produce explainable results which show both global and local patterns. Research has demonstrated that hybrid models help medical staff identify which biomarkers together with physiological signals and behavioral indicators produce the most significant impact on disease diagnosis and patient outcome prediction[111,113]. The majority of OCD studies which achieve high predictive accuracy fail to provide explainable results which makes their clinical applications restricted[84,114]. XAI in OCD serves purposes that go past technical transparency because it needs to address ethical matters and legal requirements and user needs. Research indicates AI systems can match or surpass human doctor diagnostic performance but experts continue to worry about system autonomy and patient control and privacy risks and discriminatory outcomes[70,115]. The evaluation framework FAITA-Mental Health requires systematic assessment of XAI-based mental health tools through transparency and reliability and user experience evaluation[70]. Research indicates that hybrid methods which unite DL systems with statistical and rule-based models show potential for achieving better results between prediction performance and model explainability[80,113,116]. The field now views XAI as an essential element which serves to explain model performance while enabling meaningful clinical choices and building ethical AI systems. The current research on XAI for OCD has not advanced beyond its initial development because scientists need to establish standardized explainability methods and user-oriented assessment techniques and models which must prove their effectiveness on extensive multimodal datasets[6,73,112]. The implementation of AI-supported OCD applications in clinical practice depends on addressing these essential requirements for safe and sustainable and transparent use.

CONCLUSION

This review synthesizes current evidence on the role of AI in OCD and demonstrates that AI-based approaches hold substantial potential to advance diagnostic accuracy, treatment response prediction, and personalized clinical decision support. Across studies, multimodal models integrating neuroimaging, electrophysiological, and clinical data consistently outperform single-modality approaches, underscoring the complex and multidimensional nature of OCD. Collectively, the reviewed literature suggests that AI can meaningfully contribute to the transition toward precision psychiatry in OCD when appropriately developed and clinically integrated.

The current evidence base contains multiple restrictions which block researchers from using this information for medical applications. Research studies base their findings on limited numbers of participants from individual research sites while using data that has not been verified through external or time-based assessments and showing wide variations in how researchers process information and define results. The use of untraceable “black-box” models throughout the system creates challenges for understanding model operations which damages doctor trust and blocks regulatory approval in psychiatric care where both accountability and patient protection matter most.

Research should focus on conducting extensive multicenter and longitudinal studies which will enhance model performance in real-world settings and improve their ability to generalize to different populations. The process of standardizing data acquisition methods together with feature extraction techniques and evaluation metric assessment will create conditions for successful comparison between different studies. The core design principle of XAI should be integrated into system development because it enables clinicians to understand and evaluate model-based suggestions.

The deployment of AI systems for OCD care requires human involvement through frameworks and privacy-focused learning methods and ethical governance systems. Future AI applications should operate as decision support tools which provide transparent assistance to medical professionals while they make their decisions. The future success of AI as a transformative management tool for OCD depends on solving its current methodological and ethical and translational problems.