Early-BipoLife [17] is a multicenter, prospective, naturalistic study of participants aged 15‐35 years recruited between 2015 and 2019 and assessed over a period of at least 2 years. Help-seeking youth and young adults consulting early detection centers or individuals presenting with at least one of the proposed risk factors for bipolar disorder [17,18] (refer to Section 1 in Multimedia Appendix 1 for inclusion and exclusion criteria) as well as inpatients and outpatients with depressive syndrome or ADHD were recruited at 9 sites. For detailed data collection procedures, see Pfennig et al [17], Martini et al [19], and Section 1 in Multimedia Appendix 1. Comprehensive assessments were performed after 12 and 24 months. All participants received state-of-the-art counseling and treatment. Of the 918 participants who completed the 2-year follow-up, 31 (3.17%) participants transitioned to bipolar disorder [19]. In total, 313 opted to receive a baseline MRI. Given the relatively low transition rates and the absence of a focus on predicting bipolar disorder, this sample is well-suited for predicting functional outcome using a transdiagnostic approach.

Following clinical assessments were included in the analysis of functional outcome: age, sex, medication (none or any current psychotropic medication), contact to mental health services present or past, migration status (positive when person itself, father or mother were not born in Germany and as negative if all 3 were born in Germany), first-degree relatives for bipolar disorder, DSM-IV (Diagnostic and Statistical Manual of Mental Disorders [Fourth Edition]) diagnoses and substance use (SKID-I [Structured Clinical Interview for DSM-IV Axis I Disorders]) [20], depressive symptoms (Inventory for Depressive Symptomatology–Clinician-Rated [IDS-C]) [21], early life stress (Childhood Trauma Questionnaire [CTQ]) [22], GAF [10] at baseline and in the past year, Functioning Assessment Short Test (FAST) [23,24], prodromal psychotic symptoms (Prodromal Questionnaire-16 [PQ-16]) [25], bipolar risk factors (EPIbipolar) [18], extended Bipolar-at-Risk criteria (BARS) [26], and Bipolar Prodrome Symptom Scale-Prospective (BPSS-P) [27]. In total, 126 items were included as features (for details, refer to Table S1 in Multimedia Appendix 1). All features were included at the item level, except for PQ-16, which was only available as a total scale score in the study database.

Although more differentiated measures of functional outcome do exist, we used GAF, as it is highly established and largely present in longitudinal datasets (including Early-BipoLife), which enables larger training and validation samples. Rather than predicting GAF as a continuous variable, we dichotomized GAF into impaired and nonimpaired for the following reasons: (1) to include trajectories that remained stable on a low functioning level, as well as those that impaired during the follow-ups, (2) include both follow-ups in one outcome variable, and (3) reflect a hypothetical clinical scenario, where typically a binary decision should be made between assigning and not assigning to an intervention. We defined the negative outcome as GAF equal to or below 60 (ie, moderate symptoms such as flat affect and circumstantial speech, occasional panic attacks, or moderate difficulty in social, occupational, or school functioning, such as few friends, conflicts with peers or coworkers) at least 1 follow-up. This value is in alignment with previously used cutoffs in the psychosis risk literature (Koutsouleris et al [12]: Social and Global Functioning: Role scales ≤7; Lalousis et al [28]: Global Assessment of Functioning-Symptom [GAF-S] score of ≤60). Clinically, functional impairment at this level may be considered as an indication for clinical (outpatient or inpatient) intervention, as compared to GAF 70‐61, which denotes good functioning with only mild symptoms and difficulties.

In post hoc analyses, we aimed to predict the following additional outcomes: (1) higher and lower GAF cutoff values (ie, GAF ≤55 and ≤65), (2) we used time series k-means clustering implemented in Scikit-learn v. 1.5.2 [29] to cluster the GAF trajectories in 2 clusters using 6 time points (GAF values at past year to baseline, baseline, past year to 1-year follow-up, 1-year follow-up, past year to 2-year follow-up, and 2-year follow-up) and differentiate between participants belonging to these clusters. (3) Maximum GAF values over the past year at follow-up 1 or 2, aiming to capture long-term impairments that may have been overlooked by the GAF value at follow-up assessments.

We acquired 1 mm isotropic resolution structural T1-weighted images using Siemens Magnetom MR scanners at 6 sites (Trio, Skyra, and Prisma models) and a Philips Achieva scanner at 1 site. We standardized the pulse sequence parameters across all sites to the extent permitted by each platform. For a detailed description of the scanning protocol, including the details of MRI scanners, specific hardware configurations, and pulse sequence parameters, see Vogelbacher et al [30].

Prior to preprocessing, we performed the data acquisition and quality assessment using the MRIQC tool [31]. Two authors visually inspected the obtained reports of several metrics, including a movement plot and a plot of the background noise. In this way, 23 participants were excluded from further analysis due to strong movement (n=18), ghosting (n=1), or fold-over artifacts (n=4).

We preprocessed the T1-weighted structural MRI using Freesurfer 6.0 (Martinos Center for Biomedical Imaging; cortical and subcortical parcellations) and 7.1.1/7.2.0 (Martinos Center for Biomedical Imaging; hippocampal subfields and amygdala nuclei), partially running on the Center for Information Services and High-Performance Computing (ZIH) by TU Dresden [32]. We obtained regional cortical thicknesses and surface area values for 68 cortical brain areas defined by the Desikan-Killiany atlas [33], 14 volumes of subcortical structures [34], 18 amygdala nuclei, 24 volumes of hippocampal subfields [35,36], that is, 124 MRI features in total (refer to Section 2 in Multimedia Appendix 1 for additional details). We performed a standardized quality control of the cortical and subcortical segmentations according to the established protocols of the Enhancing NeuroImaging Genetics through Meta-Analysis (ENIGMA) working group [37]. This included a visual inspection of the segmented regions using the internal and external surface methods, as well as statistical outlier detection. The outliers were subjected to further visual inspection. For details on preprocessing and exclusion of participants who did not pass the quality control or displayed major segmentation errors (n=8), please see our previous publications [38,39].

First, we predicted the impaired status at follow-up using a linear support vector machine (SVM; LIBSVM 3.1.2; developed by Chih-Chung Chang and Chih-Jen Lin [40], C-SVC, L1-loss) with instance weighting, implemented in the Neurominer Toolbox v. 1.2 (Nikolaos Koutsouleris) [41] using baseline clinical features in 357 participants after excluding those with missing GAF values at any follow-up and further removing participants with missing values for 25% or more of the features.

Second, we trained separate SVM classifiers using baseline clinical and MRI features in 124 participants with available baseline MRI. We then applied stacked generalization, where the decision scores from both base models were combined using another linear SVM that used the decision scores from the completed analyses as input features to predict the outcome group [12].

All pipelines included feature-wise scaling (range 0 to 1) and imputation of missing values using the median of the 7 nearest neighbors (Hamming distance for nominal features with 2 or fewer unique values, Euclidean distance for ordinal features). Covariates sex, age, and intracranial volume were regressed out using partial correlations exclusively in the MRI feature set, as sex and age were considered informative in the clinical domain. Both the removal of nuisance covariates and the imputation of missing data were carried out separately for test and training sets. The imputed values were derived from the training test and applied to the test set.

We used a nested CV scheme, with an outer loop (CV2) using 6 folds (leave-one-site-out, where each site represented a study site) and an inner loop (CV1) using 5 folds for model selection and C parameter tuning. In each inner fold, we applied a wrapper-based greedy feature selection approach to retain the ratio of selected features to total participants at 1:5. As a result, k=71 features were selected for the clinical models and k=25 for the MRI and stacking models. The regularization parameter (C) was optimized across 9 values ranging from 0.00390625 to 256. In the SVM analyses, probability estimation was not enabled. This means the classifier did not use a probabilistic cutoff (such as 0.5) but relied on the standard SVM decision function with a fixed decision boundary.

We assessed the contribution of features toward the classification using the sign-based consistency, which was defined as the consistency of the weights assigned by the model to a given feature, reduced by the fraction of models that removed that feature during the wrapper-based feature selection. Sign-based consistency of 1 means that the weights of the feature all share the same sign and the feature has been selected by all classifiers in the inner CV loop, whereas a value of 0 means that positive and negative weights occurred equally or the given feature was omitted in all CV folds [42].

In order to improve clinical interpretability, we analyzed the most predictive set of features retrieved by the SVM analysis (defined as sign-based consistency score P<.05) using a decision tree classifier implemented in Scikit-learn (v. 1.5.2). Briefly, the decision tree minimizes the entropy by splitting the data so that each subset contains as few mixed class labels as possible (ie, each leaf ideally contains a single class). For a binary classification, entropy=1 for a leaf composed of 50% of class 1% and 50% of class 2, whereas entropy =0 for a leaf containing only a single class. Using a 5-fold CV, we trained and evaluated the model using the following parameters: decision criterion=entropy, maximal depth=3 layers (for interpretability reasons, we decided against more layers), minimal sample splits 2‐10, class weighting to account for imbalanced classes. Class predictions were based on the default scikit-learn behavior, in which each leaf node assigns the class with the highest empirical probability (equivalent to an implicit 0.5 threshold in binary classification).

To assess clinical utility, we performed decision curve analysis, calculating net benefit across threshold probabilities and comparing each model with treat-all and treat-none strategies [43].

We used a locally hosted Llama-3 (llama-3.3-70b-instruct-q4km) via the llama.cpp framework on a local hospital computer [6]. For each participant, we converted the clinical assessments into reverse-engineered text notes by substituting item’s scores by their exact descriptions from the assessment instruments (eg, if the participant scored 0 in the item falling asleep of the IDS-C, the text note included the sentence: “I never take longer than 30 min to fall asleep.” Refer to Section 3A in Multimedia Appendix 1 for an example). This process was performed automatically using a Python script. We performed no further processing of text notes. All predictions using the Llama-3.3-70B-Instruct-q4km model were generated using deterministic decoding parameters to ensure reproducibility. We set the temperature to 0, top_p to 1.0, and max_tokens to 256. A fixed random seed was applied for all inference calls. Under these settings, model outputs were identical across repeated runs. The following items were included: age, gender, diagnoses, substance use, IDS-C, CTQ, GAF baseline and past year, FAST, information on past and present psychiatric treatment, and first-degree relatives for bipolar disorder. Llama-3 was instructed using a prompt to classify the functional outcome group (Section 3B in Multimedia Appendix 1). To further investigate the generalizability of the prompting approach, we repeated the analysis in an external sample of 590 participants aged 15‐35 years from the FOR2107 longitudinal sample of persons with affective disorder [44] (refer to Table S2 in Multimedia Appendix 1 for the details of the validation sample). The reverse-engineered notes included age, gender, main diagnosis, HAMD (Hamilton Depression Rating Scale), CTQ items, and GAF baseline.

The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2013. All procedures involving human subjects or patients were approved by the Ethics Committee of the Medical Faculty at the Technical University Dresden (no. EK290082014) and all local ethics committees. We obtained written informed consent after providing comprehensive information about the study aims and procedures.

Overall, the global functioning improved within the 2-year follow-up (mean_baseline 61.6, SD 15.7; mean_year1 69.2, SD 14.3; mean_year2 70.6, SD 14.4). Refer to Figure 1 for the trajectories.

The participants impaired at follow-up fulfilled significantly more frequently criteria for psychiatric diagnoses, particularly affective, anxiety, and substance use disorders (Table 1). They also more frequently contacted mental health services. The nonimpaired group contained more first-degree relatives of persons with bipolar disorder.

The impairments in global functioning at baseline were more pronounced in the impaired group in all domains and items of the FAST inventory (Table S3 in Multimedia Appendix 1). For distributions of GAF scores at follow-ups, refer to Figure S1 in Multimedia Appendix 1.

Participants not included in the analysis due to missing data or failed quality assessment with available GAF at baseline (n=855) displayed better clinical functioning at baseline (mean GAF_discarded 64.77, SD 18.87 vs GAF_included 61.58, SD 15.73; P=.005). Participants who opted for MRI differed neither in their baseline characteristics nor in the functional impairment outcome compared to those who did not (Table S4 in Multimedia Appendix 1). For the breakdown of demographic characteristics to single sites, refer to Table S5 in Multimedia Appendix 1.

The linear SVM classifier trained on clinical measures alone (sample size, n=357) to predict the impaired status within the 2-year follow-up achieved balanced accuracy 69.2%, sensitivity 64.3%, and specificity 71.9% (to assess classifier performance across the full range of thresholds, see the receiver operating characteristic curve in Figure S2 in Multimedia Appendix 1). Training the classifier using MRI features (sample size n=124) exclusively achieved balanced accuracy 54%, sensitivity 43.5%, and specificity 60.3%. In the same subsample, using only clinical features achieved a balanced accuracy of 73% (sensitivity 65.2% and specificity 80.8%). Combining the clinical and the MRI classifiers via stacking did not improve the performance (balanced accuracy 65.6%, sensitivity 54.3%, and specificity 76.9%).

The most consistently predictive features were items related to occupational functioning, interpersonal relationships, cognitive functioning, psychotic and affective symptoms, as well as the presence of anxiety disorder (Table 2; Figure 2; and Figure S3 in Multimedia Appendix 1). As GAF baseline was retrieved as the most predictive item, we trained an SVM classifier exclusively using GAF baseline as a single feature, achieving a comparable performance (balanced accuracy 70.2%, sensitivity 74.8%, and specificity 65.6%). However, for clinical decision-making, reliable probability estimates are essential, and single-feature classifiers typically yield poor calibration. To evaluate both the accuracy and reliability of probability estimates, we constructed calibration curves for single-feature and multifeature models. The multifeature model achieved a lower Brier score (0.214 vs 0.311), indicating more accurate and reliable probability estimates (Section 4 in Multimedia Appendix 1).

Training the decision tree using the set of most consistently predictive features (Table 2) retrieved a significant decision tree with the following parameters: balanced accuracy 76.6%, sensitivity 78.3%, and specificity 74.9%. The stratification using baseline GAF was further refined by combining the sequence with the best GAF in the past year and items related to social functioning, early life adversity, and psychosis screening. For the visualization of the decision tree, refer to Figure 3. Training the decision tree using GAF baseline only achieved balanced accuracy 72.7%, sensitivity 67.4%, and specificity 78%.

Classification using full-text notes yielded balanced accuracy 72.6%, sensitivity 78.3%, and specificity 66.9%. After excluding GAF baseline value, Llama-3 achieved balanced accuracy 69.7%, sensitivity 88.4%, and specificity 51%. In the external sample of persons with affective disorders (Table S2 in Multimedia Appendix 1), LLM achieved balanced accuracy 61.2%, sensitivity 55.4%, and specificity 70%. After excluding GAF, Llama-3 achieved balanced accuracy 60.8%, sensitivity 46.9%, and specificity 74.7%.

Decreasing GAF thresholds for impaired outcome (ie, increasing the impairment) revealed improved balanced accuracies for MRI data after reducing the feature set to hippocampus and amygdala subregions (Table S4 in Multimedia Appendix 1). However, stacking amygdala and hippocampus features with clinical data revealed no improvement (73.0% vs 72.3%, respectively). Data-driven clustering of GAF trajectories using 6 GAF time points into 2 clusters revealed a low- and high-functioning cluster (Figure S4 in Multimedia Appendix 1). However, MRI data were not able to differentiate between the 2 clusters (Table 3). The decision curve analysis for the SVM, decision tree, and LLM showed that neither the decision tree nor the SVM or LLM provided a positive net benefit compared with a “treat none” strategy (Figure S5 in Multimedia Appendix 1). In order to elucidate the relationships between the most predictive variables and functional outcome, we performed partial dependence plot analyses (Figure S6 in Multimedia Appendix 1).

SVM regression to predict the maximum GAF values over the past year to follow up using clinical features achieved a significant positive relationship between predicted and actual values (R²=0.12; P<.001) with a mean absolute error (MAE)=10.0 GAF points for clinical data. MRI features revealed a weak relationship (R²=0.03; P<.04, MAE=11.3) and there was no relevant improvement using stacking (R²=0.14; P<.001, MAE=9.8). Repeating the primary SVM analysis to predict impaired outcome using different machine learning methods did not improve the balanced accuracies for clinical and structural MRI features (Table S6 in Multimedia Appendix 1).