The data for this study were derived from the e-Prevention study database, which included 38 patients with psychotic spectrum disorders (schizophrenia, schizoaffective, and bipolar disorder). All patients were sequentially referred by their treating psychiatrists in remission, and a diagnostic evaluation was performed at the intake interview by 2 trained psychiatrists using DSM-5 criteria. At the intake interview, we recorded demographic data including age, gender, marital status, birthplace, occupation, educational level, and BMI. Patients were screened for a history of medical, neurological, psychiatric, and developmental disorders; family psychiatric history; birth complications; and history of substance abuse (smoking, alcohol, cannabis, and other psychotropic substances) (Table 1). Verbal IQ scores were derived from the administration of the Vocabulary Test of Wechsler Intelligence Scales [32] (Table 1).

For each patient, we recorded the duration of the psychotic disorder before entering the study for each patient and medications including antipsychotics, antidepressants, mood stabilizers, and benzodiazepines (Table 1). Antipsychotic medication dosage was transformed into chlorpromazine equivalents [33]. Antidepressant medication dosage was transformed into fluoxetine equivalents [34]. Finally, benzodiazepine dosage was transformed into diazepam equivalents [35].

Each participant provided written informed consent for participation in the study and monitoring of clinical and digital phenotypes after being given a full description of the study design and objectives. Moreover, participants provided a separate consent form allowing the use of their personal data for the purposes of the e-Prevention research project in compliance with the General Data Protection Regulation (GDPR, EU Regulation 2016/679). The e-Prevention project study protocol and participant consent forms were reviewed and approved by the ethics committee of the University Mental Health, Neurosciences and Precision Medicine Research Institute “COSTAS STEFANIS,” (aprroval number: 778/14-5-2019). No compensation was provided to participants.

For the total duration of their participation in the e-Prevention, patients agreed to visit the Psychiatry Department of Eginitio University Hospital once every month for an evaluation during which psychopathology (PANSS) [4], BMI, and medication status (including equivalent transformed dosage of antipsychotics, antidepressants, and benzodiazepines) were assessed, among other clinical parameters [29]. The dimensional categorization of the PANSS scores [6] was used to derive the following 5D scores: (1) positive dimension, (2) negative dimension, (3) cognitive/disorganization dimension, (4) depression/anxiety dimension, and (5) excitement/hostility dimension.

The e-Prevention database includes digital phenotypes collected passively from patients via a commercial smartwatch (Samsung Gear S3). Patients agreed to wear the smartwatch for the total duration of the study, continuously day and night, excluding charging time and activities that might lead to hardware failure (taking a shower or swimming). Details about system architecture, smartwatch feature extraction, and preprocessing are described in detail in our previous reports [30,31,36,37].

The accelerometer and gyroscope 3D signals were used to derive the short-time energy (STE) of their Euclidean norm over each 5-minute time window using the following formula:

where x is the norm of the respective accelerometer or gyroscope signal, j is the index of each sample under summation, and n is the total number of samples within the time window. The short-time energy can be thought of as the traditional measure of total signal energy (ie, the sum of the squared signal values), which can intuitively be seen as a quantitative measure of activity across the window which is used for summing. This is close to the interpretation of the activity counts of the widely used ActiGraph devices, without using thresholding for getting the counts and then summing. For each 5-minute interval, we derived a short-time energy value from the accelerometer, which measures accelerometer motor activity, and from the gyroscope, which measures gyroscope motor activity.

Heart measurements via photoplethysmography included beats per minute and intervals between successive pulses (interpulse intervals). Interpulse intervals were processed for removal of artifacts, and the count of the preprocessed interpulse intervals, named normal-to-normal intervals, for each 5-minute window was derived. This measure of normalized heart rate is considered a proxy of sympathovagal balance, or the ratio of sympathetic to parasympathetic activity [38]. The series of interpulse interval values (in milliseconds) was used to calculate the root mean square of their successive differences for each 5-minute interval. This measure of heart interpulse variation is considered a proxy of parasympathetic heart activity [39].

The sleep:wake ratio per day and the mean number of total steps per minute for each day as a measure of locomotive motor activity were also recorded. To improve the robustness of these measurements, we considered data only for days for which we have 18 or more hours of recorded data.

Our dataset included 740 patient months (minimum 1 month, maximum 26 months; mean: 11.8 months for each of the 38 patients; Table 1). We excluded months for which fewer than 720 five-minute measurements of wake data or fewer than 480 five-minute measurements of sleep data were collected. For each patient, we calculated the monthly mean and SD for each digital phenotype, separately for periods of wakefulness and sleep (except for sleep:wake ratio and locomotive motor activity). These data are summarized in Table S1 in Multimedia Appendix 1, alongside data from clinical interviews regarding psychopathology or medication. We derived the following digital phenotypes for each participant for every month: (1) mean and SD of accelerometer motor activity during wakefulness and during sleep, (2) mean and SD of gyroscope motor activity during wakefulness and during sleep, (3) mean and SD of normalized heart rate during wakefulness and during sleep, (4) mean and SD of heart interpulse variation during wakefulness and during sleep, (5) mean and SD of sleep:wake ratio, and (6) mean and SD of locomotive activity.

The results of the first step, exploratory analysis for the relation of digital phenotypes to the PANSS dimensions using random intercept LME models, are presented in Table S2 in Multimedia Appendix 1. Table 2 presents the results of the application of the LME model selection for selected models that survived step 1 analysis.

Table 2 presents the results for the best-fitting LME model analysis selected after the 3-step process described in the methods section. Selected models of digital phenotypes predicting each PANSS dimension score are grouped for each PANSS dimension. In each model, we present the estimated coefficient for the fixed effect of the digital phenotype and its corresponding SE, as well as the t value and degrees of freedom (df), the P value, and the effect size (f²) for this fixed effect. In the last 3 columns, we present the X² and corresponding P value for the comparison of the −2 log-likelihood of the different models at each step of the analysis (methods) and the best-fitting model.

In our LME models, increases in positive symptom dimension scores were related to increases in mean normalized heart rate during wakefulness and during sleep (Table 2). LME models including autocorrelation as well as random slope effects had significantly better fits than the simple random intercept-only models. Using these models, the relation of normalized heart rate during wakefulness and sleep to the PANSS positive dimension became nonsignificant (Table 3). Additionally, increasing positive symptom dimension scores were related to decreasing mean heart interpulse variation during sleep. More complex LME models (autocorrelation as well as random slope effects) did not lead to improved fit (Table 2), so the simpler model was retained.

Increases in negative symptom dimension scores were related to decreases in accelerometer motor activity (mean and SD), as well as decreases in mean locomotive activity and mean gyroscope motor activity during wakefulness. Table 3 presents the results of the application of the LME model selection. The more complex LME models had significantly better fits than the original random intercept models in all cases. Using these models, all the relations that were selected in the first-step analysis remained significant (Table 2).

Across repeated monthly assessments, higher cognitive/disorganization scores were associated with lower monthly variability (SD) of normalized heart rate during wakefulness. The more complex LME models did not provide a significantly better fit compared to the simpler step 1 model.

Increasing depression/anxiety symptom dimension scores were related to increasing mean accelerometer and gyroscope motor activity as well as their respective monthly variability (SD) during sleep (Table 3). The more complex step 2 and step 3 LME models were not significantly better compared to the step 1 model, as shown in Table 2.

Lastly, increasing excitement/hostility symptom dimension scores were associated with increasing mean gyroscope motor activity and increasing variability (SD) of accelerometer and gyroscope motor activity during sleep. There was also a relation between this symptom dimension and increasing mean normalized heart rate during wakefulness and during sleep, as well as a decreasing mean sleep:wake ratio (Table 2). The more complex step 2 and step 3 LME models were not significantly better compared to the step 1 models, except for the mean normalized heart rate during wakefulness. In that case, the more complex LME model (random slope) had a significantly better fit and was retained (Table 2). In this model, the effect of normalized heart rate during wakefulness on the excitement/hostility dimension score was not significant.

The results of the first-step LME model analysis for the effects of demographic, clinical, medication, and time variables on the digital phenotypes and on psychopathology dimensions are presented in Tables S3 and S4, respectively, in Multimedia Appendix 1, while significant results with an effect size >0.01 are further elaborated in Tables 3 and 4.

Table 3 presents the results of linear mixed model analyses for the significant effects of demographic and clinical factors on digital phenotypes that also had an effect size >0.01 (selected from Table S3 in Multimedia Appendix 1). The LME model estimated mean and SE for each value of the factor are presented in column 2, the t or F value (factors with 3 levels) for the fixed effect and its corresponding degrees of freedom (df) are presented in column 3, and finally, the P value for the fixed effect and its effect size are presented in columns 4 and 5.

Table 4 presents the results of the selected linear mixed model analyses of the significant effects of antipsychotic medication and time on digital phenotypes that also had an effect size >0.01 (selected from Table S2 in Multimedia Appendix 1). The LME model estimated mean and SE for the antipsychotic medication (chlorpromazine equivalents) and time (months) are presented in column 2, and the t value for the fixed effect and its corresponding degrees of freedom (df) are presented in column 3. The P value for the fixed effect is presented in column 4, and its effect size f² in column 5.

Table 5 presents the results of linear mixed model analyses for the significant effects of demographic and clinical factors on psychopathology dimensions that also had an effect size >0.01 (selected from Table S3 in Multimedia Appendix 1). The LME model estimated mean and SE for each value of the factor are presented in column 2, the t or F value (factors with 3 levels) for the fixed effect and its corresponding degrees of freedom (df) are presented in column 3, and finally, the P value for the fixed effect is presented in column 4. All t tests were 2-tailed.

Table 6 presents the results of the selected linear mixed model analyses of the significant effects of antipsychotic medication and time on psychopathology dimensions (selected from Table S3 in Multimedia Appendix 1) with effect sizes >0.01. The LME model estimated mean and SE for the antipsychotic medication (chlorpromazine equivalents) and time (months) are presented in column 2, and the t value for the fixed effect and its corresponding degrees of freedom (df) are presented in column 3. The P value for the fixed effect is presented in column 4, and its effect size in column 5.

We then investigated potential confounding effects of demographic, clinical, medication, and time variables on the relations of digital phenotypes with PANSS dimensions (see Methods). To be eligible for inclusion in this analysis, a potential confounder had to be significant in predicting both the digital phenotype (Table S3 in Multimedia Appendix 1) and the PANSS dimension (Table S4 in Multimedia Appendix 1).

To begin with, the effect of mean normalized heart rate during wakefulness on the positive dimension was tested with antipsychotic medication as a confounding covariate. The interaction LME model analysis showed a nonsignificant interaction of mean normalized heart rate during wakefulness and antipsychotic medication on the PANSS positive dimension (t_95.6=−1.86, P=.07) and separate significant effects of antipsychotic medication (t_98.3=2.56, P=.01) and mean normalized heart rate during wakefulness (t₄₈=2.32, P=.02). The effect of mean normalized heart rate during sleep on the PANSS positive dimension was tested with antipsychotic medication as a covariate. This LME model analysis showed a nonsignificant interaction of mean normalized heart rate during sleep and antipsychotic medication on the PANSS positive dimension (t_302.9=1.08, P=.28) and nonsignificant effects of antipsychotic medication (t_297.8=−.18, P=.85) and mean normalized heart rate during sleep (t_43.5=1.2, P=.23). Finally, the effect of mean normalized heart rate during wakefulness on the positive dimension was tested with time of follow-up as a confounding covariate. The interaction LME model analysis showed a nonsignificant interaction of mean normalized heart rate during wakefulness and time on PANSS positive dimension (t_502.1=−1.3, P=.2). The effect of time was not significant (t_505.4=−0.94, P=.35), while the effect of mean normalized heart rate during wakefulness was significant (t_539.8=2.71, P=.007).

The effect of the variability of accelerometer motor activity during wakefulness on the negative dimension of PANSS was tested with antipsychotic medication as a covariate. This LME model analysis showed a significant negative interaction of the variability of accelerometer motor activity during wakefulness and antipsychotic medication on PANSS negative dimension (t_231.6=−3.15, P=.002) and a significant effect of antipsychotic medication on PANSS score (t_365.3=3.98, P<.001) while the effect of the variability of accelerometer motor activity during wakefulness on PANSS negative dimension was not significant anymore (t_49.4=−0.84, P=.4). The effect of mean gyroscope motor activity during wakefulness on the negative dimensions of PANSS was tested with gender as a cofactor. This LME model analysis showed a nonsignificant interaction of mean gyroscope motor activity during wakefulness and gender on PANSS negative dimension (t₃₁=1.1, P=.28), while the effect of gender on PANSS score was significant (t_65.9=2.85, P=.006), and the effect of mean gyroscope motor activity during wakefulness on PANSS negative dimension was also significant (t_34.1=−2.95, P=.006). Finally, the effect of mean locomotive activity on the negative dimension of PANSS was tested with antipsychotic medication as a covariate. This LME model analysis showed a significant negative interaction of mean locomotive activity and antipsychotic medication on PANSS negative dimension (t_93.1=−2.08, P=.04) and a significant effect of antipsychotic medication on PANSS negative dimension (t_271.2=2.67, P=.008) while the effect of mean locomotive activity on PANSS negative dimension was not significant (t_34.8=−0.88, P=.38). All t tests were 2-tailed.

In this study, we investigated the potential relation of continuously recorded digital phenotypes to the monthly variation of psychotic symptom dimensions measured with PANSS. We observed significant relations between psychotic symptomatology and digital phenotypes by applying the dimensional model composed of 5 symptom dimensions (positive, negative, cognitive/disorganization, depression/anxiety, and excitement/hostility). Notably, each symptom dimension was associated with a different set of digital markers, providing further evidence for the applicability of the dimensional model.

Positive and excitement/hostility dimension symptoms were related to increased heart rate during wakefulness and sleep, indexing sympathovagal imbalance toward sympathetic activation, although only the increase of heart rate during sleep to increased excitement/hostility during sleep remained significant after application of more complex LME models including different slopes for this relation across participants. In our preliminary report [29], we showed that increased heart rate in wakefulness and sleep was related to the PANSS positive symptom scale score. Present results corroborate these initial findings in the sense that items from the PANSS original positive scale score are included in these 2 symptom dimensions (positive and excitement/hostility). A novel finding of this study was the specific relation of the positive symptom dimension to decreased heart interpulse variation during sleep, reflecting decreased parasympathetic activity, which was notably not shared with any other symptom dimension (Figure 1). The lack of universality of this relation across symptom dimensions is of particular interest given the vast body of literature ascertaining a connection between decreased vagal tone and schizophrenia and other psychotic disorders ([41-44], see [45,46] for a review). It is conceivable that decreased parasympathetic activation could be linked solely to the existence or exacerbation of positive symptomatology rather than the entire range of symptoms characterizing psychotic disorders, although this hypothesis requires further testing. From a neurobiological standpoint, this relation could be explained in association with the Neurovisceral Integration Model [47], in conjunction with the well-established dopamine hypothesis [48]. In short, cortical disinhibition (mostly in the prefrontal cortex) mediated by decreased vagal tone could subsequently lead to subcortical disinhibition, which could in turn result in dopaminergic hyperactivity and thus exacerbation of positive symptoms.

Increased negative dimension scores were related to decreased motility during wakefulness overall, marked by a decrease in accelerometer and gyroscope motor activity, a decrease in the variability of accelerometer motor activity, and a decrease in locomotive activity (Figure 1). A decrease in motility would be the expected overt behavior accompanying increased negative symptomatology [49,50]. It should be noted that in our preliminary report, we did not detect a relation of motility to the negative symptom scale score of PANSS. In that report, we observed that increased original PANSS negative scale scores were linked to decreased heart interpulse variation during wakefulness (indicating decreased parasympathetic activity) and increased sleep:wake ratio, findings which are in line with the literature cited above. While these effects were also detected in our study (Table S2 in Multimedia Appendix 1), they were excluded from further analysis due to their very small effect size (Cohen f²<0.01). This discrepancy between studies could be attributed to the different definitions of the original negative PANSS scale and the negative symptom dimension. Another plausible explanation could be that negative symptoms evolve much slower in time, and a longer period of observation would be needed to detect a robust relation to digital phenotypes measuring behavior.

Regarding the cognitive/disorganization symptom dimension, a significant, negative relation with the monthly variability of normalized heart rate during wakefulness was observed (Figure 1). This result could be indicative of disrupted sympathovagal balance related to cognitive/disorganization symptoms. In a related review [39], no studies examining heart-related autonomic dysregulation in mental processes such as attention or memory are cited. There are, however, 2 studies [51,52] in which social functioning is linked to a lower LF/HF ratio, indicative of an imbalance in autonomic regulation (higher sympathetic activity or lower parasympathetic activity). While results are not directly comparable due to differences in both the independent and dependent variables used, the alignment concerning autonomic dysregulation in this broader area of symptoms warrants further investigation.

The 2 mood-related dimensions (depression/anxiety and excitement/hostility) were related to increased gyroscope motor activity and its variability, as well as elevated accelerometer motor activity during sleep. The depression/anxiety dimension was also related to a higher variability of accelerometer motor activity during sleep. These findings are in line with established clinical observations, namely that increased motility is indicative of restless sleep, which in turn is characteristic of both depression and manic episodes [53-55]. Finally, the excitement/hostility dimension showed a unique negative relation to sleep:wake ratio, confirming that these symptoms are accompanied by less sleep (Figure 1).

The main objective of this study was to investigate the relations of digital phenotypes to psychopathology dimensions as well as demographic, clinical, medication, and time variables. Importantly, it could be hypothesized that these relations are confounded by demographic, clinical, medication, and time factors. If that were true, there should exist a direct connection among digital phenotypes and demographic, clinical, medication, and time variables, and among psychopathology dimensions and demographic, clinical, medication, and time variables simultaneously. While demographic variables were related to digital phenotypes (mainly gyroscope motor activity and heart interpulse variation), very few relations were detected between clinical variables and digital phenotypes. To conclude, the observed relations were generally not confounded by demographic or clinical factors. Finally, there were very few effects of time of follow-up on digital phenotypes.

Another finding with significant implications is that only 2 digital phenotypes varied significantly with diagnosis. The SD of gyroscope motor activity during wakefulness was larger for patients diagnosed with bipolar disorder compared to those with schizophrenia, and the monthly SD of heart interpulse variation during sleep was larger for patients with schizophrenia compared to those with bipolar disorder. If one compares these findings with the results concerning the relation of psychopathology dimensions to digital phenotypes, it is evident that digital phenotypes were more successful at dissociating the different psychopathology dimensions of psychotic disorders than the categorically defined syndromes of schizophrenia and bipolar disorder. These results further cement the dimensional model of psychosis [56].

As far as interaction effects among the 3 sets of variables (digital phenotypes; demographic, medication, and time factors; and psychopathology dimensions) are concerned, the only ones involving all sets of variables were those including antipsychotic medication or time confounders. Considering the significance of the link between normalized heart rate and the positive symptom dimension, we highlight that there was no significant interaction between normalized heart rate and either antipsychotic medication or time in predicting positive symptom dimension, conclusively ruling out antipsychotic medication and time as potential confounders. The same cannot be stated for the relation between decreased motility and negative symptoms. Significant interaction effects indicate that the observed reduction in motility is in part mediated by the effect of increased antipsychotic medication.

From a clinical perspective, our results suggest that digital phenotyping could provide a complementary framework to traditional assessment via infrequent interviews, helping clinicians move toward dynamic, dimension-specific monitoring. Our approach can enrich existing, fully streamlined platforms such as mindLAMP [57], an application which collects active data (via surveys, questionnaires, etc) and passive data from a smartphone (GPS, accelerometer, call and text logs, etc). The application can be used for research purposes, while data can also be viewed by clinicians to plan interventions accordingly, assuming the patient provides consent. Smartwatches used in this study also offer bioinformation-rich heart rate data for real-time analyses, while requiring minimal effort from the patient, thus not hindering their everyday life and activities. Furthermore, current algorithms predict impending relapse with a significant but not reliable accuracy, leaving the possibility of false positives, which complicates clinical decision-making. The implementation of the dimensional approach also entails error; however, it offers flexibility to the clinician. By monitoring digital markers and obtaining predictions for specific symptom changes, a clinician can incorporate guidelines and experience on a case-by-case basis. For example, observing a monthly average change in heart rate of 60 would imply a change of 5 units in the PANSS positive dimension, potentially implicating persistent stress. The clinician could then adjust the treatment regimen, notify the patient with the intent of scheduling a visit, or continue to observe the digital marker trajectory. Although this effect size is small, further analyses incorporating more data could refine the models and produce far stronger symptom estimates. In summary, the differentiation of digital markers across psychopathology dimensions implies that clinicians could eventually tailor interventions based not only on observed symptomatology during infrequent visits, but also on objective, continuous data streams reflecting a patient’s everyday functioning.

Another finding with significant clinical implications is that symptom dimensions were more strongly associated with objective markers compared to static diagnostic labels. Given that symptom overlap is common among psychiatric disorders, it may be sensible to use longitudinal sensor-derived data as “digital biomarkers” to stratify patients not only by diagnosis but also by predominant symptom domains and their temporal dynamics. This could support more rational pharmacological decision-making, aid in evaluating treatment response in real-world settings, and provide an objective complement to subjective self-report scales during maintenance treatment. Over time, and with the accumulation of large-scale, real-world datasets, predictive models could be refined, enabling fully personalized clinical decisions, aligning with the ongoing efforts in precision medicine.

One potential limitation of this study is the small to medium sample size of patients. We believe that this limitation is countered by the long observation time, which allowed for a substantial amount of data collected for each patient. This resulted in a large database of digital phenotype and psychopathological estimation data (about 700 data points) that increased the power for detecting significant effects. That being said, the sample was relatively homogeneous, consisting of Greek outpatients aged 18‐45 years in the maintenance phase, which could compromise external validity. To address this, further studies are necessary across different regions and involving patients at various stages of the disorder.

Another limitation of this study lies in the implementation of a coarse-grained approach using LME modeling in our analysis. Although digital phenotypes were measured continuously in time, data from the entire month preceding each psychopathology measurement were collapsed into a single mean and variance value for each phenotype. While this approach serves to produce robust and interpretable results, which we valued highly in this study, it does have drawbacks. First, it is not sensitive to the short-term (ie, days or weeks) temporal evolution of digital phenotypes and thus cannot capture its potential significance in predicting psychopathology. Second, the LME approach cannot account for nonlinear effects and is not well suited to explore all the large number of total possible interaction effects. Third, LMEs can use specified covariance matrices [eg, AR(1)] but cannot fully capture complex autocorrelations which may be present in longitudinal datasets. Finally, monthly aggregation and LME application assume missingness at random, introducing bias in the case that missingness affects predictions. To address some of these concerns, we performed a basic sensitivity analysis, aiming to test our LME model approach against random forest and Gaussian Process Regression models, which can capture nonlinear relations. Both machine-learning approaches yielded modest explanatory power (R²<0.25 across PANSS dimensions) and similar patterns of predictor importance (Table S5 in Multimedia Appendix 1). These results suggest that potential nonlinear effects do not materially change the observed associations and support the interpretability-oriented LME modeling framework adopted here.

This study adds to the mounting evidence in favor of the 5-factor model of PANSS, which can capture symptom profile heterogeneity, as well as symptom variation in time. Crucially, each psychopathology dimension was associated with a different set of digital markers, while those relations were not confounded by demographic or clinical variables. This finding constitutes a preliminary indication that biological changes such as decreased parasympathetic activity could underlie a narrower group of symptoms, rather than the entire psychotic disorder spectrum as assumed till now. Another important detail is that both behavioral markers and symptom severity vary in time, which seems inherently conflicting with static diagnostic labels and explains why only 2 of 19 used markers correlated with diagnosis. Lastly, we highlight the potential of smart devices (eg, watches or phones) and digital phenotypes not only as remote monitoring tools, but also as valuable research targets in the quest for objective markers in psychotic disorders.