Machine learning classification of ADHD and HC by multimodal serotonergic data



ADHD subjects derive from a previously reported study on SERT binding measured with [11C]DASB16.

In short, 16 patients with adult ADHD (aged 31.9 ± 10.9 standard deviation (SD), seven females) were recruited through the outpatient clinic for ADHD and affective disorders at the Department of Psychiatry and Psychotherapy, Medical University of Vienna. Twenty-two healthy control subjects (aged 33.19 ± 10.3 SD, nine females) were recruited through advertisement at the Department of Psychiatry and Psychotherapy. ADHD patients were required to be free of neuropsychiatric medication for at least three months. None of the HC were previously exposed to any psychopharmacologic treatment. All study related procedures were approved by the Ethics Committee of the Medical University of Vienna. All participants consented in written form to partake in the study after extensive explanation of the study protocol.

Subjects were screened for any somatic or neurological disorder by assessment of physical and neurological status, laboratory tests including urine drug and pregnancy tests and electrocardiography. Comorbid psychiatric disorders were assessed with the structured clinical interview for DSM-IV (SCID-I, SCID-II). Subjects with severe comorbidities or any substance abuse or addiction other than nicotine were excluded. ADHD symptomatology was evaluated by Conners’ Adult ADHD Diagnostic Interview (CAADID, Conners 1999).

Genotyping procedures

Genotyping protocols were published previously, please see for details ref.17. In summary, EDTA blood tubes of 9 ml were collected and the QiaAmp DNA blood maxi kit was applied for DNA isolation (Qiagen, Hilden, Germany). The iPLEX assay was used for genotyping on a mass spectrometer (MassARRAY MALDI-TOF). Typer 3.4 (Sequenom, San Diego, CA, USA) was utilized for genotype assignment after selection of the allele-spcific extension products. Quality control required to surpass a threshold of 80% individual and 99% SNP call rate identity of genotyped CEU trios (Coriell Institute for Medical research, Camden, NJ).

Thirty SNPs of four genetic key players of the serotonergic system, the HTR1A, HTR1B, HTR2A and TPH2 genes, were selected for this analysis based on the literature. All SNPs were coded numerically for the number of minor alleles, therfore ranging from 0 to 2. SNPs were determined based on the literature. For an overview of baseline characteristics, including genotypes (Table 1).

Table 1 Baseline characteristics with sex, age and genotypes for the total sample, HC and ADHD.

PET data acquisition

All PET and MRI scans were carried out at the Department of Biomedical Imaging and Image-guided Therapy, Division of Nuclear Medicine, Medical University of Vienna. A full-ring scanner (General Electric Medical Systems, Milwaukee, WI, USA) in 3D acquisition mode was used. For all subjects, the state-of-the-art radiotracer [11 C]DASB was used to quantify SERT binding as protocolled previously18. In summary, for tissue attenuation correction a transmission scan was obtained for five minutes with retractable 68 Ge rod sources. Data acquisition of the actual scan started with a bolus i.v.-injection of [11 C]DASB. A series of 50 consecutive time frames (12 × 5 s, 6 × 10 s, 3 × 20 s, 6 × 30 s, 4 × 1 min, 5 × 2 min, 14 × 5 min) was carried out, resulting in a measurement time of 90 min in total. FORE-ITER, an iterative filtered back-projection algorithm, was used for reconstructing the measured data in volumes of 35 transaxial sections (128 × 128 matrix). For this step, the spatial resolution was at 4.36 mm full-width at half maximum 1 cm next to the center of the field of view.

SERT quantification

The protocol for data quantification was reported previously, including preprocessing carried out in SPM12 (Wellcome Trust Centre for Neuroimaging, London, UK; In summary, the means of all time frames without visually observable head motion was used for realignment of each time frame of the dynamic PET scans. All subjects also underwent MRI scans on a 3 Tesla Philips scanner (Achieva, 3D T1 FFE weighted sequence, 0.88 mm slice thickness, 0.8 × 0.8 mm in-plane resolution). Summed PET images (integral) from realigned data was co-registered to T1-weighted images. Next, spatial normalization of the T1-weighted images was performed. Transformation of the co-registered PET images into MNI standard space was achieved by application of the obtained transformation matrices to the dynamic PET data. Finally, computation of voxel-vise images of BPND values was carried out with PMOD image analysis software, version 3.509 (PMOD Technologies Ltd., Zurich, Switzerland; and the multilinear reference tissue model with two parameters (MRTM2)19. The cerebellar grey matter without vermis and venous sinus was assigned as the reference region due to negligible availability of SERT in this region20,21.

Non-displaceable binding potential (BPND) values were extracted for regions defined according to the automated anatomical atlas (AAL). Mean values were calculated from BPND for the left and right hemispheres. Thus, a total of 49 cortical and subcortical ROI was included in the analyses.


A classification model for ADHD and HC was computed with genetic predictors, imaging predictors, all predictors as well as the top performing predictors, for each fold, respectively.

Computations were performed with the statistical software “R” ( The package “randomForest” was used for application of the eponymous algorithm (RF)22,23. In short, RF is an ensemble tree classification tool that randomly selects subsamples of observations and builds a decision tree for optimal splitting of these observations according to an outcome variable by a combination of predictors. For each split, the best performing predictor out of a random selection is applied. Generally, a higher number of predictors allowed for selection leads to optimal splits but also low diversity of the individual trees. Therefore, restricting the number of features can generate models that perform worse in the training set but are more flexible when exposed to new data. Here, 3000 trees were grown (ntree = 3000) for each model to enable multiple predictions for all patients. Classification was performed with a five-fold cross-validation (CV) design to allow optimal validation in absence of an independent test set24. If hyperparameters must be tuned, nested CV is the gold-standard technique to prevent data leakage from training to the validation phase. For RF only the number of features randomly selected at each split (mtry) can be optimized; however, there is a standard of using the square root of the number of predictors. To prevent overfitting, no optimization of mtry was performed for this analysis.

For variable selection, a combination of established algorithms “Boruta” and “varSelRF” for “R” were used23,25. Comparable to permutation-based importance evaluations, “Boruta” doubles the predictors included in the model by generating “shadow predictors” that show randomly interchanged values for each observation. Then 500 iterations of RF are run and only those predictors performing better than the best “shadow predictor” by a p-value threshold of 0.01 are preserved. These relevant predictors were then included in a backwards variable elimination algorithm, “varSelRF”. The best performing combination of predictors was then applied to the test set corresponding to each fold of the CV.

The whole CV procedure was repeated ten times and average accuracy is reported. See also Fig. 1 for a synopsis of the CV design.

Fig. 1: Graphical representation of the five-fold CV design.

CV was performed with standard settings for variables randomly selected for splitting at each node (mtry = square root of the number of predictors) and variable selection based on imputation testing as provided by “Boruta” and backwards feature elimination as provided by “VarSelRF”. The top perfroming predictors of each training set were used for classification of the respective test set. The whole CV procedure was repeated ten times and results were averaged. CV cross-validation.

There is no established method of power calculation for RF. Research indicated stable predictive capabilities of RF and comparable machine learning algorithms when enough observations are available, even in high dimensional data with the number of variables surpassing that of observations26,27. For this dataset, a ratio of 79 predictors to 38 subjects was observed.

In addition to the results produced by RF, a mixed model was computed with the “lme” package for “R”. Linear mixed regression models for BPND with diagnosis, ROI and the most informative genetic predictors included as fixed effects and subject as random effect were built. Main and interaction effects (up to three-way) were computed. Mixed model results were corrected for the number of tests and models with a corrected threshold of p ≤ 0.001. Based on these results, logistic regression models for each ROI and SNP with diagnosis as outcome variable and the respective ROI/SNP interaction term were computed. Logistic regression results were not corrected.

Source link


Please enter your comment!
Please enter your name here