Development and internal validation of a depression severity prediction model for tinnitus patients based on questionnaire responses and socio-demographics

Tinnitus is a complex condition that is associated with major psychological and economic impairments – partly through various comorbidities such as depression. Understanding the interaction between tinnitus and depression may thus improve either symptom cluster’s prevention, diagnosis and treatment. In this study, we developed and validated a machine learning model to predict depression severity after outpatient therapy (T1) based on variables obtained before therapy (T0). 1,490 patients with chronic tinnitus (comorbid major depressive disorder: 52.2%) who completed a 7-day multimodal treatment encompassing tinnitus-specific components, cognitive behavioural therapy, physiotherapy and informational counselling were included. 185 variables were extracted from self-report questionnaires and socio-demographic data acquired at T0. We used 11 classification methods to train models that reliably separate between subclinical and clinical depression at T1 as measured by the general depression questionnaire. To ensure highly predictive and robust classifiers, we tuned algorithm hyperparameters in a 10-fold cross-validation scheme. To reduce model complexity and improve interpretability, we wrapped model training around an incremental feature selection mechanism that retained features that contributed to model prediction. We identified a LASSO model that included all 185 features to yield highest predictive performance (AUC = 0.87 ± 0.04). Through our feature selection wrapper, we identified a LASSO model with good trade-off between predictive performance and interpretability that used only 6 features (AUC = 0.85 ± 0.05). Thus, predictive machine learning models can lead to a better understanding of depression in tinnitus patients, and contribute to the selection of suitable therapeutic strategies and concise and valid questionnaire design for patients with chronic tinnitus with or without comorbid major depressive disorder.


ADSL adsl01
"During the past week I was bothered by things that usually don't bother me." 2 ADSL adsl02 "During the past week I did not feel like eating; my appetite was poor." 3 ADSL adsl03 "During the past week I felt that I could not shake off the blues even with help from my family or friends." 4 ADSL adsl04 "During the past week I felt I was just as good as other people." 5 ADSL adsl05 "During the past week I had trouble keeping my mind on what I was doing." 6 ADSL adsl06 "During the past week I felt depressed." 7 ADSL adsl07 "During the past week I felt that everything I did was an effort." 8 ADSL adsl08 "During the past week I felt hopeful about the future." 9 ADSL adsl09 "During the past week I thought my life had been a failure." 10 ADSL adsl10 "During the past week I felt fearful." 11 ADSL adsl11 "During the past week my sleep was restless." 12 ADSL adsl12 "During the past week I was happy." 13 ADSL adsl13 "During the past week I talked less than usual." 14 ADSL adsl14 "During the past week I felt lonely." 15 ADSL adsl15 "During the past week people were unfriendly." 16  I often think about whether the noises will ever go away. 177 TQ tin44 I can imagine coping with the noises. 178 TQ tin45 The noises never 'let up'.

TQ tin46
A stronger person might be better at accepting this problem. 180 TQ tin47 I am a victim of my noises.

TQ tin48
The noises have affected my concentration.

TQ tin49
The noises are one of those problems in life you have to live with. 183 TQ tin50 Because of the noises I am unable to enjoy the radio or television. 184 TQ tin51 The noises sometimes produce a bad headache. 185 TQ tin52 I have always been a light sleeper. Overview of all 185 features which were used for classification. TQ: German version of the Tinnitus Questionnaire [5]; PSQ: Perceived Stress Questionnaire [3]; SF8: Short Form 8 Health Survey [4]; ADSL: General Depression Scale -long form [2]; SOZK: sociodemographics questionnaire; TINSKAL: visual analogue scales measuring tinnitus loudness, frequency and distress, TLQ: Tinnitus Localisation and Quality [6].
All classifiers were implemented with the statistical programming language R [17] using the package mlr [18], which provides a consistent interface to many machine learning algorithms from other R packages. A grid search was employed for hyperparameter tuning using area under the ROC curve (AUC) as evaluation measure. Table 2 provides an overview about each classifier, including used R package, tuned hyperparameters and their value ranges. All other hyperparameters were set to default values.  Supplementary C: best model The classifier that induced the best model was lasso (AUC=0.87±0.04). For each feature selection iteration, Figure 1 shows a heatmap of model reliance (M R) scores for all features. After iteration i = 1, 89 out of the 185 features were kept for model training in i = 2. The feature selection wrapper converges at i7 with a model that uses 6 features, each with a M R exceeding 1. Notably, the removal of 97% of features from i = 1 to i = 7 let to much simpler model, while only dropping marginally in AUC (-0.02). Figure 1: M R convergence for lasso classifier. Heatmap of retained and discarded features for each feature selection iteration for the lasso classifier. A blue-shaded cell represents a feature with a model reliance exceeding 1 whereas a grey cell represents a feature that was discarded from model training in the next iteration. The top subfigure shows AUC for the models of each iteration.