The inclusion and exclusion criteria are presented in Table 1.

The databases searched were Medline, Embase, PsycINFO and CINAHL. The search was conducted in May 2022 and updated in June 2024 to include new research prior to publication. Reference lists of included studies were screened to identify additional papers.

The final search strategy is presented in the Supplementary Materials. No limits were placed on the date or location of publication. An English language filter was applied to the search. The search strategy was created through an iterative process where HG, EV, SP and AK all inputted.

Results from the search were imported into Rayyan for screening (31). HG and EV independently screened the title and abstract of all papers. The interrater reliability at this stage was 0.7 (Cohen’s Kappa). Once all abstracts were screened, HG and EV undertook full-text screening of all remaining papers. HG and EV discussed disagreements in the records that were included or excluded, and in cases where a decision could not be made, SP and AK were consulted.

Key outcomes data collected were author, year of publication, country, study design and setting, assessment tool for NVB and clinical change, statistical tests and method of aggregation, type and length of intervention and outcomes and main results. Data from each study were extracted into an Excel spreadsheet. Three studies were subject to an initial data extraction by HG and EV and compared to confirm that the result was satisfactorily similar. The interrater agreement for the data extraction process was 83%. HG and EV discussed any discrepancies before proceeding with the data extraction for the remaining studies independently. The full data extraction form is presented in the Supplementary Materials.

The CASP cohort study checklist was used to assess the risk of bias in each included paper (32). Three studies were subject to an initial quality assessment pilot by HG and EV and were compared to confirm that the result was similar. HG and EV extracted data from the remaining papers independently. Disagreements were to be resolved by separate discussions with SP and AK.

Due to the heterogeneity of methods used to measure and record NVB, formal synthesis and meta-analysis of results were not possible. Results were subject to narrative synthesis by summarising the main outcomes of each paper and determining which changes in NVB were significantly associated with clinical improvement or deterioration. We compared the behavioural profiles for depression and schizophrenia, and whether these differed across intervention types. We also explored the overall quality and strengths and weaknesses of the evidence and highlighted key knowledge gaps and avenues for future research.

We identified 9–743 publications from the searching stage and screened 7–493 titles and abstracts after removal of duplicates. Following abstract screening, 75 studies underwent full text screening, and 20 studies met the final eligibility criteria for inclusion (Figure 1). Table 2 summarises the properties of included studies. 15 studies observed patients with depression, two studies observed patients with schizophrenia and three studies observed separate groups of patients with depression and patients with schizophrenia within the same study.

Study sample sizes ranged from 12–164 participants, with the average being 43 participants. Interventions included pharmacological intervention, psychological intervention, neuromodulatory intervention, and a variety of non-standardised interventions including hospitalisation, pharmacological and psychological interventions. Interventions ranged in length from 2–32 weeks, with the average length being 12 weeks. Most studies (14/20) assessed patient outcomes using The Hamilton Rating Scale for Depression (HRSD).

Regarding the analysis of NVBs, 12/20 studies used manual analysis, 6/20 used automated analysis, and 2 studies compared both automated and manual methods. NVB was assessed most frequently during a videotaped clinical interview based on the administration of the HRSD. Other assessment settings included videotaped psychotherapy sessions or less structured mental state examinations.

Most studies (15/20) used group comparison tests or correlations to compare NVBs before and after treatment, showing treatment induced changes in NVBs. Five studies used multilevel modelling or regression approaches, and two studies extended the analysis to make predictive inferences about how NVBs may predict clinical change. Fiquer et al. (2013) (38) found that adaptive gestures (self-touching and object touching) at baseline predicted higher negative affect scores after treatment (p=0.04). Fisher et al. (41) found that an increase in emotional valence during the therapy session predicted later reductions in depression symptom scores (p=0.018).

Heterogeneity across studies made synthesis challenging, mostly due to the wide variety of behavioural categorisation and measurement approaches used. For example, some studies focused on measuring specific facial AUs and discrete movements, whereas others grouped behaviours into broader categories and ascribed emotional states (such as happiness, anger and restlessness) to patterns of NVBs behaviours. This led us to focus our synthesis on broader categories of behaviour. A wide variety of methods to measure NVBs and lack of standardised procedure also impaired comparisons. Some studies measured NVB continuously throughout the clinical interview, others for just a short portion of the interview, and the features of NVB that were recorded varied across studies (e.g. duration, frequency and intensity). It was also challenging to determine if automated methods led to more consistent results due to the still broad range of approaches used to measure and categorise behaviour among this subset of studies. Among the studies employing automated coding methods, NVBs measured included nonverbal synchrony between the patient and interviewer (34, 49), emotional valence (41), changes in AUs (16, 43, 50), changes in behaviours associated with specific emotions (44), behavioural depression degree (45). This emphasises the need for more standardised measurement approaches to strengthen conclusions.

Definitions of key NVBs discussed in this section are presented in Table 3.

The synthesis of results firstly focused on the analysis of behaviours relating to smaller, discrete parts of the face (e.g., eye and eyebrow movements, mouth movements) and the relationship between discrete behaviours and responses to treatment. Results were highly variable, and the only consistently occurring facial behaviour associated with symptomatic improvement in patients with depression was increased mouth movements/smiling, observed in 4/18 studies (16, 30, 43, 47). Only two studies assessed facial behaviours in patients with schizophrenia, therefore a robust conclusion cannot be drawn, although both studies both observed a significant reduction in facial activity and expressiveness over time (42, 50).

Due to the lack of consistency across specific facial behaviours, we focused our analysis on more general behaviours such as body movements, gesturing, emotional expressivity, and emotional valence. Studies seemed to shift from more specific methods of analysing NVB in older studies (e.g., using FACS or ethograms comprised of specific muscle movements) to a more global and generalised approach of assessing the nature of NVB changes in studies published from 2021 onwards. A more generalised approach may enable inter-individual differences in specific facial movements and differences in methods of measuring NVB to be accounted for. 16/20 studies conducted a more global assessment of behaviours (as opposed to assessing facial AUs only or blinking/eye movements), 14 studies included patients with depression in the sample, and 2 included both patients with depression and patients with schizophrenia in the sample.

Generally, as clinical improvement occurred over time, patients with depression showed increasing levels of social and affiliative behaviours, including head movements (5/14 studies) (16, 35–37, 47), gesturing (2/14 studies) (37, 47), body contact/touching (35–37) (3/14 studies), facial expressivity and positive emotional valence (6/14 studies) (38, 41, 44, 45, 47, 51), movement (including restlessness and fidgeting) (3/14 studies) (35, 40, 47). Such behaviours portray an image of the patient transitioning from behaviours associated with social withdrawal and diminished emotions to more active and engaged as treatment progressed. Although increased levels of restlessness and fidgeting could also be interpreted as occurring due to discomfort. Hence, despite the differing definitions of body movement and restlessness, many studies did not differentiate between the two behaviours nor acknowledge that restlessness and fidgeting could be associated with negative valence.

Conflicting results were reported by Bouhuys and Albersnagel (1992) (35) and Fiquer et al. (2017) (39) who found that patients who did not improve showed more speaking-effort behaviours which contrasts with the theory of pro-social behaviours being associated with improvement. In addition, Gaebel and Wölwer (42) (2004) reported that as symptoms improved, facial expressivity and emotional repertoire reduced.

No consistently occurring behaviours were observed across the two studies which assessed more general behaviours in patients with schizophrenia (30, 42).

7 of the 16 studies that measured generalised behaviours assessed the association between NVB and pharmacological intervention (16, 35–37, 39, 42, 51). This subset of studies corroborated with some of the overall findings which showed an increase in body touching (35), gesturing (51), nodding and head movements (16, 35–37), and a general increase in expressivity and AUs (16, 51) as patients received pharmacological intervention. One conflicting result in this subset of studies was that of facial expressivity, Ulrich and Harms (1985) found increased facial expressivity as symptoms improved whereas Gaebel and Wölwer (2004) showed reduced facial expressivity after 3 and 4 weeks of antidepressant treatment respectively.

3 studies measured the effect of neuromodulatory treatments (2 deep brain stimulation (44, 47), 1 transcranial direct current stimulation (38)) and found that patients showed increased nodding and eye contact after 2 weeks of transcranial direct current stimulation and increased facial expressivity (38), more reactive, engaging and fidgeting behaviours (including head movements, laughing, smiling, general mouth movements, raised eyebrows, head touching and illustrative gestures), and fewer behaviours associated with disengagement after 6–8 months of deep brain stimulation (44, 47).

3 studies observed NVB changes whilst patients underwent courses of psychotherapy, and found increased positive emotional valence (41), increased expression entropy (45), and increased nonverbal synchrony (49) as treatment progressed.

In 12/20 studies, patients commenced medication or treatment at the same time of the baseline interview (30, 35–39, 41, 44–48). 2/20 studies included patients who were treatment naïve or medicated at the baseline interview (42, 50), and the remaining 6/20 studies did not specify if patients were already being treated prior to or commencing treatment at the time of the baseline interview (16, 34, 40, 43, 49, 51).

5 studies acknowledged and discussed the impact of medication effects on the study results. 1/5 studied the effect of medication on NVB in patients with depression, 1 included a sample of both patients with depression and patients with schizophrenia, and 3 explored the effect of medication on NVB in patients with schizophrenia (35, 42, 46, 48, 50).

In studies observing NVBs in patients with depression, Bouhuys et al. (1992) (35) mentioned that the pattern of behaviours at baseline did not significantly differ between medicated and unmedicated groups, but acknowledged that treatment conditions in the study were not controlled for which may have impacted findings.

Mackintosh et al. (1983) (46) studied a combined sample of patients with schizophrenia and patients with depression and found that there were no significant differences in the blink rate when patients were compared based on medication (patients were receiving one of three types of treatment: minor tranquilisers, tricyclic antidepressants, or major tranquilisers combined with antidepressants).

The studies exploring medication effects in schizophrenia described the impact on results in more depth. Meya and Renfordt (1986) (48) compared the effect of haloperidol and perazine in previously unmedicated patients with schizophrenia, and found that patients treated with haloperidol showed a tendency towards increased eye contact with the interviewer during the first 14 days of treatment compared to patients treated with perazine. The authors stated that more research is required to validate this finding and explore why this may have occurred.

Gaebel and Wölwer (2004) (42) conducted exploratory analysis of differential treatment effects of haloperidol and perazine in patients with schizophrenia, and found a significant medication effect regarding the facial repertoire and facial activity in the acute phase of schizophrenia, stating that an attenuation of facial expressivity between baseline and T1 was more pronounced in patients treated with haloperidol than perazine, and more pronounced in patients treated with perazine versus unmedicated control participants. When comparing unmedicated versus medicated patients at baseline, there was no significant difference in facial expressivity. The authors discussed that attenuated facial activity seen in the study is unlikely to be due to drug treatment and more likely to be explained by habituation to the interview, stating that the effect size was too small to be of clinical significance. Gaebel and Wölwer’s (2004) study did not align with the results observed by Meya and Renfordt (1986), signifying that the question of the influence of haloperidol and perazaine on alterations in NVB remains unanswered.

Further, Schneider et al. (1992) (50), included one group of patients (group 1) with schizophrenia who were unmedicated at the start of the study, and group 2 who commenced the study whilst continuing to take neuroleptic medication. The study found a significant reduction of facial action over time in group 1, and no significant change in group 2, despite there being no significant difference at baseline between groups in psychopathology scores. The authors also stated that more research is required to determine if reduced activity induced by neuroleptics is due to both motor and emotional sedation. The lack of consistent conclusions across all three studies represents a significant gap in the literature on the effect of medication on NVB in schizophrenia and the subsequent relationship to symptom scores and patient outcomes.

Results from the risk of bias assessment using the CASP Cohort Study checklist (32) are presented in Table 4. Regarding areas of concern, just 9/20 studies clearly identified all confounding factors, 8/20 studies accurately measured the exposure (e.g. interview condition) and outcomes (clinical outcome and NVBs) to minimise bias, and only 5/20 studies had results that could be applied to the local population due to small sample sizes.

Strengths of the evidence base included that all studies addressed a clearly focused issue, 15/20 studies had a follow up of subjects that was sufficiently complete, and 13/20 studies followed up subjects for long enough.

Sources of bias were largely related to the study design. Many studies did not consider confounding factors – e.g., if raters were blinded, if the same interviewer was used throughout, if the interviewer was familiar to the patient, which medications and dose were prescribed, and other patient demographics or comorbidities which may influence behaviours. The impact of cultural differences could not be extensively discussed due to limited reporting of patient demographics and limited acknowledgement of the impact of cultural factors within the included studies.

We also noted how the use of a variety of outcome measures when evaluating clinical improvement may account for the differences between studies. For example, across studies that used the HRSD to measure clinical outcomes we observed discrepancies regarding thresholds and questionnaire versions.

More is also to be understood about how NVBs change throughout the interview process. For example, reductions in NVB may occur due to habituation to the interview process (42). Several studies mitigated this by only recording NVBs during the first part of the interview (36, 38, 47). Still, studies that used this method differed in the length of the initial segment of the interview that was coded, ranging from 5–20 minutes.

Participants will also have increased familiarity with the interviewer throughout subsequent sessions which may impact NVB and introduces variation and uncertainty whereby patients within the same study with differing levels of familiarity with the interviewer are compared. Bhatia et al. (2019) acknowledged that one potential reason for not observing increased head movement synchrony between the patient and interviewer as clinical improvement occurred could be due to a lack of consistency in whether or not patients saw the same therapist during the 21 week trial – some patients had the same therapist, others may have changed therapist up to 4 times. This phenomenon has been recorded in the wider literature whereby the interviewer’s NVBs and nonverbal synchrony may influence patient improvement (18, 53).

Such sources of bias are implicit to using a clinical interview as the environment in which NVBs are measured. The benefits of assessing behaviours during a clinical interview are that it enables the observation of a larger range of behaviours that would occur during a social interaction which may not occur during an image viewing task or another means of assessing behaviour, however it is important to note that results from such studies may be influenced by the interpersonal relationship between the patient and the interviewer.

Areas of potential bias, such as expectation bias from the NVB raters if they are aware of the patient’s clinical status, can be overcome with the increasing use of automated means of NVB analysis. The use of automated analysis in future studies may also help to study the potential habituation to the interview by coding behaviours occurring across the entire interview, as it can be inferred that a reason for only analysing segments of interviews is due to the time burden associated with manual coding.

General increases in movement and expressivity seen over time in patients with depression is in line with the Social Withdrawal hypothesis of depression, which suggests that patients with high levels of depression show less social or affiliative behaviours to maintain or increase interpersonal distance (16); a response that may have biological and evolutionary roots to protect the self from failure and/or harm (54–56). These observations could translate into a biomarker for assessing treatment response in depression.

There were studies outside the scope of this review which supported these findings. Two studies of NVBs in remitted patients with depression revealed that patients who went on to relapse displayed less active listening behaviours during the discharge interview compared to patients who remitted, and patients who showed less involvement behaviours at baseline had less favourable outcomes (57, 58).

Results which conflicted with this hypothesis were reported by Bouhuys and Albersnagel (1992) (35) and Fiquer et al. (2017) (39) who saw more speaking effort and restlessness in patients who did not improve. A potential explanation was that when patients with high speaking effort evoke excessive social support, they may be more prone to interpersonal distress and depression maintenance and persistence (39). Gaebel and Wölwer (42) (2004) found that patients had reduced expressivity as symptoms improved and that facial behaviours stabilised over a 3 month period. The authors stated that the reason for NVBs not changing as clinical improvement occurs could be due to habituation to the interview setting or a mismatch between the experience of emotion and displaying the emotion via facial expressions. Another explanation is offered by Fisher et al. (41) (2023) who found that emotional expressions follow a U shaped curve whereby positive emotions reduce as patients face challenging emotions in the middle of a course of therapy, but that the level of positive emotions at the end of treatment is higher than at the beginning.

One study outside of the scope of this review compared NVBs between patients with depression and a healthy control group. At baseline, the patient group showed higher levels of negative NVB (frowning, crying, head and lips down), and lower levels of positive NVB (eye contact and smiling). However, after eight weeks of treatment the patients with depression had no significant changes in their NVBs regardless of clinical improvement. The authors suggested that NVB may be independent of depression severity and may rather be a sign of depression (59).

We identified several noteworthy papers that investigated how NVB can be used to predict clinical change prospectively or retrospectively but could not be included in our review as NVB was only measured at one time point (57, 58, 60–62). Two papers studied behavioural predictors of amitriptyline response in depressed inpatients (60, 61). One showed that patients who did not respond to medication showed more body-focused movements, postural changes and reduced smiling at the baseline interview (61). Another found that non-responders demonstrated more assertive and affiliative behaviours (defined as a behaviour with an attractive function to draw the attention of the interviewer to the individual, and that express communicative efforts for social interaction) and had more engagement with the interviewer, compared to no significant behaviours associated with treatment responders (60). These findings from the wider literature contrast with results from this review in which clinical improvement in depression is associated with affiliative and interactive NVBs. This further exemplifies the need for larger scale studies and longer term follow ups that span the entire course of treatment.

There were no consistently occurring behaviours identified across studies in patients with schizophrenia. Results were inconclusive from so few studies thus it cannot be determined if NVBs are a useful metric to measure treatment response in schizophrenia. Further research is needed to discern the impact of medication on NVBs and whether behaviour changes are a result of medication side effects or symptom improvement. There is a need to identify patients at risk of deterioration, and conversely, early identification of patients who may be improving, to guide treatment strategy. Treatment resistance is linked with a higher risk of clinical deterioration, hospitalisation, poor quality of life and impaired real-world functioning, therefore the early identification of patients who fail to respond to initial interventions can improve the treatment approach at an earlier phase and contribute to improving outcomes.

The paucity of research in this area is exemplified by the fact that no studies since 2004 have been conducted to further explore the role of NVBs in monitoring treatment response over time in schizophrenia. Specifically, the lack of studies on clinical deterioration could allude to potential reporting or publication bias, in that many of the NVB studies assessed in this review were conducted in parallel with an interventional clinical trial, and only studies which saw clinical improvement in patients (rather than deterioration) may have ultimately sought publication. In addition, it may be more challenging to recruit patients into long-term studies who could be more acutely unwell or at risk of deterioration.

NVBs could be a potential avenue on the search for reliable signals that allude to early treatment response, however the literature, in its current format, does not support this use.

There are two important factors which may have a strong influence over NVBs in patients with schizophrenia: the impact of negative symptoms and the effects of medication. Patients with higher levels of negative symptoms typically show reduced levels of NVB expression and a greater avoidance of social contact (63, 64). Avolition and blunted affect can impede social perception in patients with schizophrenia and ultimately reduce NVB expression (65). Furthermore, the presence of negative symptoms may be associated with a higher rate of spontaneous movement disorders in patients prior to any exposure to antipsychotic medication which may further affect NVB production (66). Movement disorders such as spontaneous dyskinesia or parkinsonism may therefore be related to the pathophysiology of schizophrenia in a proportion of patients rather than a side effect of antipsychotics, though medication may also alter the expression of NVBs (67, 68).

Known side effects of antipsychotic medication include reduced ability to move the eyes, facial activity, blink rate, and emotional bluntness (46, 50, 69, 70). Two studies included in the review discussed the potential impact of antipsychotic medication on facial expressions and reported conflicting results. One concluded that attenuated facial expressivity in patients with schizophrenia over time is unlikely to be due to medication effects, is unlikely to be clinically significant and is more likely to be attributable to habituation to the interview process (42). Conversely, Schneider et al. concluded that the reduction in facial expression is likely to be due to medication effects. Patients who were already medicated at baseline showed no change in their facial expression over time. Patients who were unmedicated at baseline but commenced medication during the study showed reduced facial expression over time, despite no significant difference in the psychopathology scores between groups. This finding raises questions about whether the observed reduction in movement is a result of high levels of agitation at baseline or due to side effects of antipsychotic medication and EPSEs (50). It is important to note the small sample sizes in either study (n=16 and n=34), exemplifying the importance of further research. A potential use of the further study of NVB in schizophrenia could be in trials comparing different antipsychotic regimes to determine the level of impact of the medication on NVBs and emotional blunting, and how alleviating this could support the expression of prosocial behaviours and social functioning, and facilitate recovery (17).

The study design may also influence the lack of consistent findings across patients with schizophrenia. Patients with different subtypes were included together in the analysis but considering different manifestations of schizophrenia and experiences of positive and negative symptoms, this may have not been appropriate. Compared to depression, the field of schizophrenia is far less researched and despite the implicit challenges outlined above, there is considerable scope in the evaluation of NVBs as predictors of clinical outcomes or change. This could be demonstrated in a large-scale, long-term study considering confounding factors (medication type and dosage, medical history, positive and negative symptoms) that have not been sufficiently accounted for in the existing literature to confirm or dispute historic findings. This is especially important considering that the last papers exploring NVBs in patients with schizophrenia were published 20 years ago so do not reflect recent treatment advancements.

To our knowledge, this is the first systematic review to examine NVB indicators of treatment response in adults with SMI. We performed a robust systematic search in line with PRISMA guidelines. Limitations of the review relate to the limited evidence and lack of large-scale studies, meaning the results must be interpreted with caution when generalising to the wider patient population. Additionally, requiring studies to record NVB at multiple time points or only during a clinical interview may have excluded other relevant papers that provide predictive insight into how NVBs relate to clinical improvement or deterioration.

Due to the heterogeneity of methods used to measure NVB’s, specific definitions used to define behaviours or groups of behaviours may vary across studies whereby the same behaviour is measured differently, which may be a limitation in our interpretation and analysis of behavioural results.

Technologies that can automatically analyse facial expressions and emotions from video provide opportunities to incorporate real-time NVB analysis into routine video consultations or smartphone-based monitoring software (71), particularly given the increase in remotely delivered mental health care and the growing market for mental health apps (72). This integration of NVB analysis into routine mental health assessment and treatment processes could be beneficial to its broader adoption and consistent use, given that engagement with mental health apps is typically low and attrition is high, particularly if such technologies do not integrate seamlessly into mental health services (73). Having a behavioural biomarker that can be assessed in real time (such as NVBs being monitored during a routine consultation) may be preferable to the need for additional apps and technologies that collect data outside of clinical consultations (72). This may also reduce the burden of data collection and monitoring for both patients and healthcare professionals. This review concluded that there are several behavioural changes that could indicate clinical improvement in depression (head movements, quantity of movements, smiling). These are fairly simple, objective and easy-to-recognise behavioural markers that if validated further, could reasonably be tracked and monitored in real time during appointments.

Telemedicine and remote patient monitoring have led to statistically significant improvements in health outcomes, healthcare costs, patient satisfaction and accessibility across many disease areas (74). There is not yet a large evidence base on outcomes from the use of novel mental health technologies (75), and much more research is needed to address unknowns regarding the longer-term consequences and potential unintended consequences of digital monitoring interventions (72). Yet the fact that innovative monitoring technologies have been successfully integrated into care pathways in other chronic disease areas shows that enhancing mental health services with advanced digital solutions and novel behavioural biomarkers is a realistic possibility (73).

It is important to emphasise the potential challenges of assessing NVBs in psychiatric disorders. Firstly, the NVBs associated with a particular mental illness may vary across individuals within the same group, making it difficult to develop standardised assessments¹⁸. Additionally, the presence of mental, physical, and/or substance use comorbidities can further complicate the assessment of NVB (76–79). Similarly, treatment can affect the expression of NVBs, often not allowing to separate the effects of the illness from medication side effects. Therefore, future research should consider controlling for comorbidities, effects of medication and individual differences.

To improve the accuracy and reliability of NVBs as predictors for treatment response, future studies may benefit from standardising and controlling confounding variables and clinician and rater’s blindness to patient’s clinical or treatment status. Due to the sample sizes being small for most of the studies, future research would benefit from observing changes in NVBs on a larger scale of patients with a range of symptom severities and longer follow up periods, with strictly controlled treatment regimens and a control group where possible. We also noted a lack of studies on clinical deterioration. Designing future studies to focus on potential markers of early deterioration and relapse, rather than just clinical improvement, is an important area of further research and will further enable us to understand the utility of NVB biomarkers in either case.

Furthermore, if manual coding methods are used in future studies, having multiple raters, adequate training to improve inter-rater reliability and ensuring that interviews are consistent in timings and context may improve the reliability and validity of results. However, consideration should be given to the use of AI-based technology and automated methods to detect and analyse NVBs, reducing the potential of human rater bias (80) and enhancing the speed and efficiency of analysis. Incorporating AI-based technology could also help to enable remote monitoring and larger scale, cross-cultural studies taking place (81). At the same time, it is critically important to establish the reliability, validity and ethical considerations of AI-based technology and the collection of sensitive behavioural data (82, 83). Given the novelty of the field, data protection and ethical frameworks for the collection of NVB data and development of AI-based tools are currently lacking but essential for the sustainable widespread use of such technologies that preserve patient autonomy (72, 75).