Trained mental health professionals used a semistructured clinical interview—the Kiddie Schedule for Affective Disorders and Schizophrenia (K-SADS)—to screen for and establish diagnoses in participants [18]. The K-SADS is designed to establish psychiatric diagnoses in youth between the ages of 6 and 18 years.

Trained mental health professionals assessed the clinical severity of OCD using the Children’s Yale-Brown Obsessive Compulsive Scale (CY-BOCS) [19]. The CY-BOCS interview begins with a checklist of obsessions and compulsions to establish which symptoms have been present over the past week. On a scale from 0 to 4, clinicians rate 5 items on the severity of obsessions and 5 items on the severity of compulsions. Severity is rated on 5 dimensions: level of distress caused by the symptoms, functional interference, time consumed by symptoms, and resistance to and degree of control over symptoms. A total severity score is calculated by summing all 10 items on a scale from 0 to 40. Although CY-BOCS scores are continuous, a previous study of 815 youth between the ages of 4 and 18 years found the following cutoff scores to be consistent with global clinical severity ratings: 0-7, subclinical; 8-15, mild; 16-24, moderate; and 25-40, moderate-severe to severe [20]. A CY-BOCS score of ≥16 was required to be included in the study.

The audio preprocessing system is illustrated in Figure 3. To effectively use the audio signals, the audio recordings must be segmented into shorter segments. Next, speech and nonspeech regions must be differentiated. To obtain speaker segments, the conversations will first be segmented into speech and nonspeech regions.

We will use pretrained voice activation and diarization models that will be fine-tuned with a few manually annotated ground-truth labels of speech and nonspeech samples from our data set [21]. We follow this approach to ensure a balance between resources spent in training a model while maintaining high accuracy in the obtained speaker segments. Following speaker segmentation of the audio recordings, only the audio segments corresponding to the youth are retained.

We used the vad-crdnn-libriparty pretrained model from SpeechBrain for speech detection [22] owing to its flexibility and modular structure, which is a favorable quality when applying models to clinical data owing to the tractability of errors. The pretrained model is based on a combination of convolutional and recurrent neural networks and a fully connected network. The model accepts audio segments as input and yields a posterior probability frame-level or segment-level posterior probability as output. Finally, a threshold is applied on the output posterior probability to classify the segments as speech and nonspeech.

We define a speaker segment to be an utterance of 2-4 seconds from a youth. The interviews will have a varying number of utterances. For a balanced analysis, we will randomly select 10 speaker segments per youth. Since the random selection of speaker segments may result in a high variance within the observations, we will also explore the feasibility of analyzing specific audio segments consistently over all speakers. These audio segments will be selected on the basis of insightful segments in the interview; for instance, questions in the interview associated with depression. In total, 10 interviews—5 with youth with OCD and 5 with those with no psychiatric diagnosis—have already been used to develop an appropriate method for speaker segmentation [21]. We will add the remaining 54 observations once we commence the described analysis.

From each speaker segment, we will extract the extended Geneva minimalistic acoustic parameter set [23], consisting of functionals of lower-level features. We will use the OpenSmile toolkit to extract these features and the resulting feature vector has a length of 88 [24].

The derived feature vector will be used as input to a neural network autoencoder model to obtain a latent model. The latent model is pretrained on English speech using the Interactive Emotional Dyadic Motion Capture data set and a semisupervised loss as previously described [16]. The latent model is a 2D latent space (v1 and v2) in a semisupervised denoising autoencoder with a reconstruction loss plus a loss based on the linear association between the activation labels and the latent space (the aim is a high association). Thus, to obtain v1 and v2, the 88 vocal features are projected into the latent space through the pretrained network. Previous analyses indicate that this latent space represents 2 dimensions of speech activation or intensity [16]. Sad and bored speech is characterized by low activation whereas elated and angry speech is characterized by high activation [25]. Thus, we assume that this latent space represents emotional intensity. We do not have an objective metric to evaluate this quantitatively using the current data and labels. Therefore, we will test this assumption in future work when more data labels are available.

For the machine learning analyses, we will use the raw speech signals to learn new features.

We will perform an ANOVA using a mixed effects regression model with the vocal features v1 and v2 as the outcome. We code the diagnosis, Diagnosis={0, 1} for no-OCD or OCD, respectively. The model for the jth vocal feature is as follows:

where μ_j is a constant offset for the jth endpoint, i = 1, ..., 200, y_j (youth_i) ∼ N (0, σ²_yj ), ε_i ∼ N (0, σ²_j), and the effects are mutually independent. Age, gender, and an interaction between age and gender are included as fixed effects to remove confounding effects from these variables. Diagnosis is included as a fixed effect and is our primary interest. The individual youth, denoted youth, is included as a random effect. Diagnosis is included as a fixed effect. In total, 10 repeated measures are available for each of the 64 youths included in this model.

With this analysis, we shall test the hypothesis that there is no effect of diagnosis; that is, H₀ : δ_j = 0, or equivalently that the means of the vocal features are equal for the OCD and no-OCD groups (when effects of age and gender are removed), at a 1.25% level of significance.

We will use a logistic regression with OCD severity categories moderate and severe-extreme as the outcomes. We denote the severity s = 0, 1. We will fit the following logistic regression model to the scores:

where ε_i ∼ N (0, σ_2j) and v_j is the jth latent vocal feature adjusted for age and gender and averaged over the 10 repetitions for each youth, as described in the following.

This analysis aims to test the effects from vocal features on the severity scores through the null hypotheses that b₁ = 0 and b₂ = 0, which we will test at a 1.25% level of significance.

We expect age and gender to have an influence on the vocal features and would like to remove these confounding effects from the signals. We will exclude effects from age, gender, and age and gender interactions from the vocal features through a linear regression, as follows.

First, we will find the confounding effects using a linear regression model, which estimates age and gender contributions to vocal features as follows:

where, for the ith observations and jth vocal feature, j = 1, 2. The vocal features adjusted for age and gender are then as follows:

Finally, we will use the average vocal features over the 10 repetitions for the kth youth to obtain _jk = Σiεyouth_k v_ji / 10.

We are interested in the null hypotheses associated with the effect of a diagnosis on the two dimensions in the latent model as well as the effects from the two adjusted latent dimensions on the severity scores. We will adjust our 4 P values using the Bonferroni method, and thus test each null hypothesis at a 5%/4=1.25% level of significance resulting in a family-wise error rate for the four tests of 5%.

To improve the vocal features for possible use as biomarkers of OCD severity, we shall use the following methodology. First, 2 authors (SD and NLL) will annotate samples for activation and valence on a 5-point Likert scale with 1 indicating low activation and 5 indicating high activation. We shall only provide the annotators with one speech segment at a time in a blinded manner and in random order.

For the planned analyses, we will not use the valence labels, but the activation labels as these labels have shown promise in previous investigations [16]. We will use these labels as in our previous work [16]. The major advantages of labeling speech signals with activation scores include making labels for Danish speech and for children’s speech, and labeling speech of patients and controls. Thus, the labels provide added information as no open source data exist, thus supporting transfer learning wherever feasible.

Subsequently, we shall train the following machine learning models using the following annotations: variational and denoising autoencoders with semisupervised losses based on activation labels and a reconstruction error, plus a logistic regression or logistic regression loss in a neural network to obtain a classification model.