5024

SHAP Interpretation of a Tree-Based Model for Deep Gray Matter QSM and R2* in First Episode Psychosis Patients and Their Response to Antipsychotics

Pamela Franco^1,2,3,4, Cristian Montalba^1,2,5, Raul Caulier-Cisterna⁶, Alonso González^7,8, Juan Undurraga⁹, Nicolás Crossley^1,2,7, and Cristian Tejos^1,2,10
¹Biomedical Imaging Center, Pontificia Universidad Catolica de Chile, Santiago, Chile, ²Millennium Institute for IntelligentHealthcare Engineering - iHEALTH, Pontificia Universidad Catolica de Chile, Santiago, Chile, ³Department of Physics, Faculty of Science, Universidad de Santiago de Chile, Santiago, Chile, ⁴School of of Civil Engineering, Computer Science and Telecommunications, Faculty of Engineering, Universidad Finis Terrae, Santiago, Chile, ⁵Radiology Department, School of Medicine, Pontificia Universidad Catolica de Chile, Santiago, Chile, ⁶Department of Informatics and Computing, Faculty of Engineering,, Universidad Tecnológica Metropolitana, Santiago, Chile, ⁷Department of Psychiatry, School of Medicine, Pontificia Universidad Catolica de Chile, Santiago, Chile, ⁸School of Medicine, Universidad Finis Terrae, Santiago, Chile, ⁹Department of Neurology and Psychiatry, Universidad del Desarrollo, Santiago, Chile, ¹⁰Department of Electrical Engineering, School of Engineering, Pontificia Universidad Catolica de Chile, Santiago, Chile

Synopsis

Keywords: Psychiatric Disorders, Psychiatric Disorders

Motivation: Several studies have demonstrated altered neurochemicals in psychosis. QSM quantify susceptibility changes, which have been associated with iron concentrations in dopamine pathways.

Goal(s): Identify the strongest predictive QSM and R2* in first-episode psychosis (FEP) patients and their response to antipsychotics using a tree-based model.

Approach: A tree-base model to discriminate between treatment-responsive (RS) and treatment-resistant (TRS) FEP patients by looking at tissue susceptibilities.

Results: Our model classifies RS and TRS patients: 96.67 ± 1.38 % accuracy. Also, TRS could be classified by QSM: left amygdala, right globus pallidus interna, and nucleus accumbens, which have been associated with decreased dopamine in TRS patients.

Impact: The proposed features could be used in future studies to early detect TRS-FEP patients and promptly adopt adequate treatment. This intervention may improve their clinical outcomes and minimize the functional disability and social burden resulting from prolonged psychosis.

Introduction

Several publications have demonstrated iron concentrations changes in dopamine pathways of psychosis patients (1-4). Iron concentration can be quantified measuring effective transverse relaxation rates (R2*) and quantitative susceptibility mapping (QSM) from multi-echo gradient-echo (GRE) sequences (5). Since iron is a co-factor in neurotransmitter biosynthesis, grey matter nuclei can experience changes in iron concentration, it is possible to refine disease monitoring and improve patient risk stratification (6,7). The present study aims to identify the strongest predictive QSM and R2* variables in FEP patients and their response to antipsychotics using XGBoost models.

Methods

3D multi-echo GRE and T1-weighted FLAIR of 76 first-episode psychosis patients were acquired in a 3T Philips Ingenia MRI scanner. Table 1 shows the clinical data. Twenty-four patients were classified as Treatment-Resistant Schizophrenia (TRS) according to the fulfilled TRRIP criteria for treatment resistance (8), and fifty-two patients were classified as treatment-responsive Schizophrenia (RS); they had a documented history of response to a non-clozapine antipsychotic medication. QSM and R2* were reconstructed from a 3D multi-echo GRE sequence. QSM reconstruction was performed as in (4) using Variable Sophisticated Harmonic Artifact Reduction for Phase data (vSHARP) (9) and FAst nonlinear Susceptibility Inversion (FANSI) toolbox (10,11). Images were registered and normalized to an NMI space, Figure 1. Twenty-two regions of interest (ROI) of deep GM and subcortical brain nuclei were segmented using the Multicontrast PD25 version 2019 (12,13). We calculated the mean QSM and R2* values for each ROI. A Machine Learning (ML) model was designed to select regions whose susceptibility values adequately discriminate between first-episode psychosis with and without resistance to the treatment. Furthermore, an ML was designed to select parameters that adequately separate RS and TRS patients using sequential forward selection (SFS). We used XGBoost as a classifier(14,15). The performance of the classification was evaluated using stratified 10-fold cross-validation. A confusion matrix was constructed based on prediction results in each training and validation sample. Additionally, the Pearson correlation method was used to calculate the correlation matrix between all susceptibility values. Hierarchical clustering was then applied to classify its rows/columns into groups(16,17). To identify the best predictive variables, we applied SHapley Additive exPlanations (SHAP) values to find the importance of each feature on the prediction of the model (18).

Results

Twenty-nine features were used during the modelling procedure. The model showed a predictive accuracy of 96.67 ± 1.38 % (precision: 100 ± 0%, recall: 95.37 ± 1.75%, f1-score: 97.62 ± 0.94%). The most important predictors of the model are presented in Figure 2. The most significant QSM areas were: right nucleus accumbens, left hippocampus, right globus pallidus interna, left substantia nigra, right caudate, and right subthalamic nucleus. The most significant R2* areas were left thalamus, right hippocampus, and left amygdala. The waterfall graph, Figure 3, represents the cumulative sum. Three negative values contributed to the RS class: R2* of the left thalamus and QSM of the left hippocampus and right subthalamic nucleus. While the positive values (QSM of the left amygdala and substantia nigra, right globus pallidus interna, nucleus accumbens, and caudate; R2* of left amygdala) contributed to distinguishing TRS from RS patients. Finally, an analysis of the hierarchical clustering, Figure 4, shows that the parameters selected by SHAP values corresponded to two clusters.

Conclusion

We have developed an ML approach that accurately classifies between RS and TRS patients with 96.67 ± 1.38 % accuracy. We validated how relevant these features are in the classification problem using SHAP values. SHAP-values formulation guarantees three important properties: local accuracy, missingness, and consistency (19). Therefore, its results can be interpreted as the feature's importance and reflect its influence on the prediction. Considering the relative importance of independent variables, the treeSHAP found the best nine predictive variables. Both QSM and R2* provides information about the magnetic properties of tissues, including iron content. However, from best nine predictive variables, QSM provides higher feature importance, in comparison with R2* in the evaluation of schizophrenia-related brain iron content changes to classify both classes (6). Additionally, Left Amygdala, Right globus pallidus interna, and nucelus accumbens are top three highest feature of SHAP contribution, which has been studied previously in cohorts with TRS (20-22). Therefore, the proposed features could be used in future studies to detect TRS FEP patients early and promptly adopt adequate treatment. This intervention may improve their clinical outcomes and minimize the functional disability and social burden resulting from prolonged psychosis.

Acknowledgements

This work has been funded by projects PIA-ACT192064 and ICN2021_004 of the Millennium Science Initiative Program of the National Agency for Research and Development, ANID. The authors also thank the Fondecyt projects 1181057 and 1231535 by ANID and PUENTE grant 2022-14 VRI, PUC. RC-C was funded by ANID Fondecyt Postdoctorado 2021 (Nº 3210305).

References

1.- Hare D, Ayton S, Bush A, Lei P. A delicate balance: Iron metabolism and diseases of the brain. Front Aging Neurosci. 2013 Jul 18;5:34. doi: 10.3389/fnagi.2013.00034. PMID: 23874300; PMCID: PMC3715022.

2.- Davis KL, Kahn RS, Ko G, Davidson M. Dopamine in schizophrenia: a review and reconceptualization. Am J Psychiatry. 1991 Nov;148(11):1474-86. doi: 10.1176/ajp.148.11.1474. PMID: 1681750.

3.- Kim SW, Stewart R, Park WY, et al. Latent Iron Deficiency as a Marker of Negative Symptoms in Patients with First-Episode Schizophrenia Spectrum Disorder. Nutrients. 2018 Nov 8;10(11):1707. doi: 10.3390/nu10111707. PMID: 30412998; PMCID: PMC6266210.

4.- García-Saborit M, Muñoz N, Milovic C, et al. (2023). Quantitative Susceptibility Mapping MRI in deep brain nuclei in first episode psychosis. Schizophr Bull. DOI 10.1093/schbul/sbad041.

5.- Wang Y, Liu T, Quantitative susceptibility mapping (QSM): Decoding MRI data for a tissue magnetic biomarker, Magn Reson Med. 2015; 73(1):82-101.

6.- Xu M, Guo Y, Cheng J, Xue K, et al. (2021). Brain iron assessment in patients with first-episode schizophrenia using quantitative susceptibility mapping, Neuroimage Clin. 31:102736. doi: 10.1016/j.nicl.2021.102736.

7.- Li D, Hui E, Chan Q, Yao N, Chua SE, McAlonan GM, Pang SYY, Ho SL, Mal HKF. (2018). Quantitative susceptibility mapping as an indicator of subcortical and limbic iron abnormality in Parkinson’s disease with dementia, Neuroimage Clin. 27; 20:365-373. doi: 10.1016/j.nicl.2018.07.028.

8.- Howes OD, McCutcheon R, Adig O, et al. (2017). Treatment-Resistant Schizophrenia: Treatment Response and Resistance in Psychosis (TRRIP) Working Group Consensus Guidelines on Diagnosis and Terminology. Am J Psychiatry. 1;174(3):216-229. doi: 10.1176/appi.ajp.2016.16050503.

9.- Kan, H., Arai, N., Takizawa, M., et al. (2018). Background field removal technique based on non-regularized variable kernels sophisticated harmonic artifact reduction for phase data for quantitative susceptibility mapping. Magnetic Resonance Imaging, 52, 94–101. https://doi.org/10.1016/j.mri.2018.06.006.

10.- Milovic C, Bilgic B, Zhao B, et al. (2018). Fast nonlinear susceptibility inversion with variational regularization. Magn Reson Med. 80(2):814-821. doi: 10.1002/mrm.27073.12.- Milovic C, Bilgic B, Zhao B, Langkammer C, Tejos C, Acosta-Cabronero J. (2019). Weak-harmonic regularization for quantitative susceptibility mapping. Magn Reson Med. 81(2):1399-1411. doi: 10.1002/mrm.27483.

11.- Milovic C, Prieto C, Bilgic B, et al. (2021). Comparison of parameter optimization methods for quantitative susceptibility mapping. Magn eason Med. 85(1):480-494. Doi: 10.1002/mrm.28435.

12.- Xiao Y, Bériault S, Pike GB, Collins DL. (2012). Multicontrast multi-echo FLASH MRI for targeting the subthalamic nucleus. Magn Reson Med. 30(5):627-640. doi: 10.1016/j.mri.2012.02.006.

13.- Xiao Y, Fonov V, Bériault S, et al. (2015). Multi-contrast unbiased MRI atlas of a Parkinson’s disease population. Int J Comput Assist Radiol Surg. 10(3):329-41. doi: 10.1007/s11548-014-1068-y.

14.- Xiao Y, Fonov V, Chakravarty, et al. (2017). A dataset of multi-contrast population-averaged brain MRI atlases of a Parkinson’s disease cohort. Data in Brief. 12:370-379. doi: 10.1016/j.dib.2017.04.013.

15.- Yuan KC Tsai LW, Lee KH, Cheng YW, Hsu SC, Lo YS, Chen RJ. (2020). The development of an artificial intelligence algorithm for early sepsis diagnosis in the intensive care unit. Int J Med Inform. 141:104176. doi: 10.1016/j.ijmedinf.2020.104176.

16.- Gu Y, Wang C. A Study of Hierarchical Correlation Clustering for Scientific Volume Data, Advances in Visual Computing, 2010, Volume 6455. ISBN: 978-3-642-17276-2.

17.- Ciaburro G. MATLAB for Machine Learning: Practical examples of regression, clustering and neural networks, Publisher: Packt Publishing, 2017, ISBN: 978-1788398435.

18.- Lundberg S, Lee SI. A unified approach to interpreting model predictions. Proceedings of the 31st International Conference on Neural Information Processing Systems. 2017: 4768-4777.

19.- Morde, Vishal VAS. (2019). XGBoost Algorithm: Long Way She Reign!

20.- Mier D, Schirmbeck F, Stoessel G, et al. Reduced activity and connectivity of left amygdala in patients with schizophrenia treated with clozapine or olanzapine. Eur Arch Psychiatry Clin Neurosci. 2019 Dec;269(8):931-940. doi: 10.1007/s00406-018-0965-4. Epub 2018 Dec 11. PMID: 30539230; PMCID: PMC6841919.

21.- Qi W, Wen Z, Chen J, Capichioni G, et al. Aberrant resting-state functional connectivity of the globus pallidus interna in first-episode schizophrenia. Schizophr Res. 2023 Sep 14;261:100-106. doi: 10.1016/j.schres.2023.09.018. Epub ahead of print. PMID: 37716202.

22.- Danielsson K, Stomberg R, Adermark L, et al. Differential dopamine release by psychosis-generating and non-psychosis-generating addictive substances in the nucleus accumbens and dorsomedial striatum. Transl Psychiatry. 2021 Sep 13;11(1):472. doi: 10.1038/s41398-021-01589-z. PMID: 34518523; PMCID: PMC8438030.

Figures

Table 1. Demographic and clinical data for the first-episode psychosis (FEP) patients with and without resistance to the treatment, RS, and TRS, respectively. Quantitative data are expressed as the mean (range). PANSS: Positive and Negative Syndrome Scale. * indicates statistically significant differences (p < 0.05). ** Antipsychotic doses expressed in chlorpromazine equivalents (Leucht et al., 2016).

Figure 1. Schematic description of the registration process. (a) First, we changed the resolution of the atlas PD25 version 2019 (ICBM152 space) to the MNI space. (b) Second, we co-register the T1-weight FLAIR and the magnitude multi-echo GRE. (c) Third, we normalize the T1-weight FLAIR image, and obtain the deformation field. (d) We applied the deformation field to the QSM and R2* and normalized it in the MNI space (e). Finally, we multiplied the atlas with QSM and R2* maps to obtain twenty-two mean values per map.

Figure 2. Feature importance of our models based on SHAP-values. (a) The mean absolute SHAP-values are depicted to illustrate global feature importance. (b) The local explanation summary shows the direction of the relationship between a variable and class study. Positive SHAP-values predict RS patients (dark blue color) and negative values predict TRS patients (dodgerblue color).

Figure 3. Waterfall plot explaining local disease prediction showing the contribution of each variable to the prediction. For RS vs. TRS patients from the perspective of TRS patients. Below the x-axis, the baseline value (E[f(X)]) is displayed, indicative of the expected value of the model evaluated on the background dataset. The SHAP-values of each variable are summed to match the model output with all variables included. Positive SHAP-values predict RS patients (dark blue color), and negative values predict TRS patients (dodgerblue color).

Figure 4. Dendrogram and hierarchical clustering results based on average linkage method for all quantitative susceptibility parameters of RS and TRS patients.

Proc. Intl. Soc. Mag. Reson. Med. 32 (2024)

5024

DOI: https://doi.org/10.58530/2024/5024