XGBoost-SHAP-based interpretable diagnostic framework for early cognitive impairment in type 2 diabetes mellitus

doi:10.21203/rs.3.rs-4157256/v1

Download PDF

Article

XGBoost-SHAP-based interpretable diagnostic framework for early cognitive impairment in type 2 diabetes mellitus

https://doi.org/10.21203/rs.3.rs-4157256/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Objective To develop and validate a radiomic-clinical model to assess early cognitive impairment in type 2 diabetes mellitus (T2DM) using the XGBoost algorithm.

Methods We retrospectively enrolled 193 patients with T2DM from two medical centers. According the Montreal Cognitive Assessment (MoCA), patients were categorized into normal control (NC) and mild cognitive impairment (MCI) groups. We used ComBat to normalize and gather the data distributions of two centers. The Elastic Net Regression were used to filter redundant and irrelevant features. Based on the eXtreme Gradient Boosting Machine algorithm (XGBoost), clinical factors and radiomic features was used to construct the combined model. The SHAP method explained the model by prioritizing the importance of features, in terms of assessment contribution.

Results The radscore, along with two clinical factors (education level and drinking), were used to build the combined model. The AUCs for predicting MCI in the training set, testing set, and validation set were 0.802, 0.817, and 0.852, respectively. The radscore was the most important feature for discriminating MCI/NC classification, and higher SHAP values of radscore were associated with a higher risks of MCI onset. Conversely, higher SHAP values of education level and drinking were associated with a lower risks of MCI onset. However, the contribution of drinking to the model was minimal.

Conclusion The radiomic-clinical model, utilizing the XGBoost algorithm, can be an auxiliary tool for predicting early cognitive impairment in T2DM.

Health sciences/Diseases

Health sciences/Diseases/Endocrine system and metabolic diseases

radiomics

magnetic resonance imaging

machine learning

mild cognitive impairment

type 2 diabetes mellitus

XGBoost-SHAP

• Radiomic-clinical model can be an auxiliary tool for predicting early cognitive impairment in T2DM.

• Due to its superior precision value and performance, XGBoost-based machine learning algorithms could be as a competitive alternative to regression analysis for predicting MCI.

• SHAP could explain and visualize radiomic-clinical machine learning model in a clinician-friendly way.

The incidence of type 2 diabetes mellitus (T2DM) is rising progressively with an aging population. Chronic hyperglycemia is considered a significant factor in cognitive decline.[1] People with diabetes had a 39% higher risk of Alzheimer's disease (AD) and a 47% higher risk of dementia than those without diabetes.[2; 3] Diabetes-related cognitive impairment mainly includes mild cognitive impairment (MCI) and dementia. Studies have shown that one-third of patients with MCI remain stable, one-third revert to normal cognition, and one-third develop dementia.[4] Early identification and care for individuals at the MCI stage may prevent disease progression, as many elderly individuals suffer from MCI without fulfilling the diagnostic criteria for AD.[5]

Diabetes-related cognitive impairment as a central nervous system complication is gaining attention. Researchers have focused on identifying diagnostic and therapeutic targets for diabetic cognitive impairment, using clinical biomarkers, genetic analysis, and neuroimaging techniques. Studies have shown that inflammatory markers like IL-6 and TNF-α are linked to cognitive decline. However, the predictive ability of these markers is influenced by confounding factors due to the widespread presence of inflammation in chronic diseases. β-amyloid (Aβ) deposition and Tau hyperphosphorylation are recognized as the classical pathogenic mechanisms underlying AD. Nevertheless, as an invasive examination, it is challenging for every patient to accept.[6; 7] Structural MRI and functional MRI techniques have recently emerged as effective tools for detecting cognitive decline in individuals with diabetes.[8; 9; 10; 11] However, the clinical sensitivity and specificity of the extensively studied structural MR technique focusing on the hippocampus or medial temporal lobe are relatively low.[9] Additionally, diffusion tensor imaging (DTI) and blood oxygen level-dependent functional MRI are challenging to widely use in clinical practice due to complexity and high cost.[12] Consequently, there is a pressing need for a simpler, faster, low-cost, non-invasive method to accurately identify patients with MCI.

Radiomics, as an emerging technology, can transform image data into mineable data by extracting quantitative texture features to diagnose disorders including tumors and neurodegeneration.[13; 14] This approach has the potential to uncover hidden information inaccessible through single-parameter methods like volume, providing insights into genomic, cellular, and metabolic characteristics.[13] Radiomics has been attempted in the MCI/AD field [15], but no study has investigated cognitive impairment in the T2MD population.

Therefore, the aim of this study is to establish a clinically promising method for early identification of MCI in T2DM patients by applying radiomics to whole brain regions, including gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF).

Study design and participants

The study included 165 patients with T2DM admitted to A Medical Center from January 2020 to May 2021 and 29 patients admitted to B Medical Center from January 2021 to October 2021. Patients from Medical Center A were randomly stratified into the training set (n = 116) and internal validation set (n = 49) in a 7:3 ratio. Patients from Medical Center B were included in the external validation set.

Inclusion criteria were: ① Right-handed patients aged 45 to 75; ② Patients with T2DM meeting 1999 WHO diagnostic criteria, with a diabetes duration over 0.5 years; ③ Patients capable of completing neuropsychological assessments, including the Montreal Cognitive Assessment (MoCA), Hamilton Depression Scale (HAMD), Hamilton Anxiety (HAMA), Hachinski Ischemic Score (HIS), Clinical Dementia Rating (CDR), and Activity of Daily Living (ADL);[16; 17; 18; 19] ④ Patients were conscious and cooperative during head MRI examinations, with an absence of stroke (excluding lacunar infarction) or noticeable motion artifacts in the images. Exclusion criteria were: ① Patients with acute diabetes mellitus complications such as severe ketoacidosis, hyperglycemic hyperosmolar state, and hypoglycemic coma; ② Patients with other endocrine system disorders (hyperthyroidism, hypothyroidism, pituitary dysfunction); ③ Patients with a history of alcoholism, drug addiction or substance abuse.; ④Patients with anxiety (HAMA > 14), depression (HAMD > 20), clinical dementia (CDR > 1), vascular dementia (HIS > 7), impaired daily living activities (ADL > 26), and patients with MoCA < 18 score. Patients with MoCA scores ≥ 26 constituted the normal control (NC) group, while those with scores between 18 and 25 comprised the Mild Cognitive Impairment (MCI) group. The process of including and excluding patients is illustrated in Fig. 1.

Education levels were defined based on the education system used in our country. 0 = illiterate, 1 = primary school, 2 = junior high school, 3 = high school or technical secondary school, 4 = undergraduate, 5 = graduate. Drinking history referred to consuming over 100g of alcohol per week in the past year, not meeting the criteria for alcoholism (60 g/day).[20; 21] Smoking history involved smoking one or more cigarettes per day for over 6 months.

Acquisition and analysis of MR images

MRI data was obtained using 3.0 T MRI scanner. Routine sequences included T1WI, T2WI, T2 FLAIR, and DWI. The parameters of T1WI imaging: TR = 1750ms, TE = 24m, FOV = 220× 220 cm, matrix = 256 × 256, flip angle = 111°, echo chain = 10, bandwidth = 31.25, layer thickness = 5mm, gap = 1.5mm. The parameters of T2 FLAIR: TR = 9000ms, TE = 120ms, FOV = 220 × 220 cm, matrix = 256 × 256, flip angle = 160°, echo chain = 18, bandwidth = 50, layer thickness = 5 mm, and gap = 1.5 mm. T1WI images were automatically segmented into GM, WM and CSF using the spm12 package (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/) in Matlab software. Then, Experienced neuroradiologists, blinded to clinical data, manually adjusted GM, WM, and CSF boundaries using ITK-SNAP software (http://www.itksnap.org). Modification involved: (1) removal of nonbrain tissue, brainstem, and cerebellum, and (2) adjustment of GM, WM, and CSF segmentation. T2 FLAIR imaging was used for observing and automatically segmenting WMH using the spm12 package.

Acquisition and selection of radiomics features

Whole-brain GM, WM and CSF were selected as regions of interest (ROIs). To minimize the central effect of MR images from different hospitals and scanners [37], all T1WI images underwent preprocessing. [22] Initially, all images were resampled to 1× 1× 1mm³ resolution through linear interpolation to eliminate anisotropy effects on features. Subsequently, a Gaussian filter reduced noise, and correction of magnetic field inhomogeneity helped minimize external interference effects. Finally, intensity was standardized by limiting grayscale values to a range of 0–32 to ensure unbiased comparisons. Preprocessing of images and extraction of features were performed using AK software (Artificial Intelligence Kit V3.0.0.R, GE Healthcare). Features with intraclass correlation coefficient (ICC) values exceeding 0.75 were included in the follow-up analysis to assess their reproducibility. The Mann-Whitney U test and Elastic Net Regression were then applied to filter out redundant and irrelevant features.

Radiomics model construction and validation

The model was constructed by six machine learning algorithms: LR (Logistic Regression), SVM (Support Vector Machine), Random Forest (RF), Bayes, KNN (k-nearest neighbor), and XGBoost (eXtreme Gradient Boosting Machine). The XGBoost algorithm, a decision-tree-based approach, demonstrates notable efficiency in handling missing data and aggregating weak prediction models to create highly accurate ones.[23] The SHapley Additive exPlanations (SHAP) summary plot, derived from game theory, explains the output of various machine learning models.[24] SHAP summary plots provide a visually concise representation of the range and distribution of importance that each feature has on the model's output, relating the feature's value to its impact. Features were initially sorted by their global importance. Each dot, representing the SHAP value of a feature from a patient, was horizontally plotted and vertically stacked to illustrate the density of the same SHAP value. The dots were then color-coded based on the value, ranging from low (blue) to high (red). The Area Under the Curve (AUC) from Receiver Operator Characteristic (ROC) analysis assessed the accuracy and stability of the models. The DeLong test was employed to compare the performance of different ROC curves.

Statistical Analysis

Statistical analyses were conducted using R software (version 3.5.0), SPSS software (version 17.0, Armonk, NY), and Python (version 3.5). Continuous variables were presented as (means ± standard deviations). Normality of distribution was assessed using the Kolmogorov-Smirnov test. Variables were compared using the t-test for normal distributions or the Mann-Whitney test for non-normal distributions. Categorical variables were presented as [median, interquartile interval] and compared using the chi-square test. A p < 0.05 was deemed significant.

Informed consent

to participate in the investigation was obtained in writing from all subjects and/or their legal guardian(s) in accordance with the Declaration of Helsinki.

The study proactively ensured adherence to the highest ethical standards by implementing necessary measures. It sought informed consent to prioritize participants' autonomy by outlining the study's goals and ensuring voluntary participation. Strict protocols protected confidentiality and privacy; personally identifying information was securely managed, available only to the research team, and never revealed in published data or conclusions. These ethical protections highlight the dedication to participant welfare and scientific integrity throughout the study.

Ethics approval

The study has been approved by the Ethic Committee of ZhejiangProvincial People's Hospital(ethical approval number is 2020QT174) and conducted in compliance with Good Clinical Practice guidelines and the Declaration of Helsinki. Written informed consent to participate in this study was provided by the participants’ legal guardian/next of kin.

Patient characteristics

We enrolled a total of 194 patients. In the A medical centers, there were 62 patients in the NC group and 103 patients in the MCI group, while in the B medical centers, there were 14 patients in the NC group and 15 patients in the MCI group. The scores of HAMA, HAMD, HIS, CDR, ADL, MMSE, and mod MoCA did not exhibit significant differences between the two medical centers (p < 0.05 for all; Table 1). Additionally, there were no significant differences between the two medical centers in terms of age, sex, BMI, culture, duration of diabetes, smoking, alcohol use, history of other diseases, and laboratory results (all p > 0.05; Table 2).

Performance of a clinical model

The stepwise logistic regression analysis identified education level and drinking as independent risk factors for predicting MCI (Table 3). The AUC of the Clin-model was 0.710, 0.653, and 0.793 in the training set, testing set, and validation set, respectively (Table 4).

Radiomic feature selection and performance of radiomics model

The Rad-model was constructed using all 3111 texture features extracted from the three ROIs. The AUC of Rad-model was higher than that of the other three models constructed using texture features extracted from a single ROI, as shown on the Supplementary Materials Table S1.

To mitigate the central effects of data of two medical centers, we used ComBat to normalize and gather the data distributions (Fig. 2). the AUC of Rad-model, using LR algorithm, wase 0.838, 0.733, 0.662, in the training set, inter-validation set and extra-validation set, respectively. In comparison, XGBoost algorithm demonstrated best stability and generalization (0.741, 0.722, 0.683 in training set, inter-validation set and extra-validation set, respectively), as illustrated in Table 3. The SHAP summary plot revealed that only four features significantly contributed to the predictive performance of the radiomics model in the XGBoost algorithm (Fig. 3). The wavelet.HLH_glcm_ClusterTendency_WM was the most important feature for discriminating MCI/NC classification. The coloration indicated that the model’s output increased as the SHAP value of log.sigma.3.0.mm.3D_glcm_ClusterShade_CSF decreased, which was different from others features in the cohort.

Development and visualization of radiomics-clinical combined models

The Comb-model was constructed using both radiomics and clinical risk factors. Using LR algorithms, the AUC of the Comd-LR-model was 0.861, 0.705, and 0.819 in the training set, inter-validation set, and extra-validation set, respectively. Utilizing the XGBoost algorithm, the AUC of the Comd-XGB-model was 0.802, 0.817, and 0.852 in the training set, inter-validation set, and extra-validation set, respectively (Table 5). The SHAP summary plot visually demonstrates the feature’s value affected the feature’s impact attributed to Comb-XGB-model (Fig. 4).

The DeLong test was employed to compare the two joint models. It was observed that, compared to the LR algorithm, the XGBoost algorithm sacrificed the prediction performance of the training set (0.861 vs. 0.802, p = 0.059), but the diagnostic performance of the testing and validation sets improved (0.705 vs. 0.817, p = 0.014; and 0.819 vs. 0.852, p = 0.480). Overall, the XGB model demonstrates better generalization performance than the LR model (Fig. 5, 6).

In this study, we developed a radiomics model as an adjunctive tool to identify MCI in T2DM patients, enhancing prospects for early intervention and prevention of cognitive impairment. The diagnostic model, using the XGBoost algorithm, demonstrates superior performance among various machine learning algorithms. The diagnostic performance of the model was further enhanced by integrating the clinical variables of drinking and education level.

Unlike previous radiomics studies that distinguish MCI from NC, our study primarily focused on the T2DM population. The sensitivity and specificity of our comb-XGB-model ranged from 0.710 to 0.736 and 0.767 to 0.789, respectively, resembling the diagnostic performance observed in the non-T2DM population.[15] The findings may suggest potential changes in brain microstructure between MCI and NC; however, these changes do not appear significantly different between T2MD and non-T2MD patients.

The AUCs in this study demonstrate the advantage of using an XGboost algorithm over the traditional LR classifier for early prediction of MCI probability. The SHAP method explained that only four features significantly contributed to the predictive performance of the XGBoost-based radiomics model, elucidating the reason behind observed overfitting in other machine learning algorithms. Combining radiomic and clinical analysis, we observed that the XGB algorithm significantly enhances the AUC of the testing set while maintaining a comparable AUC for the training set compared to the LR algorithm.

Our study revealed a significant positive association between educational attainment and cognitive function, which aligns with prior research findings. Our study also indicated a marginal protective impact of moderate alcohol consumption on cognitive function. Some, but not conclusive, evidence (mostly from observational studies and occasionally from clinical trials) suggests a protective association between potentially moderate alcohol and cognitive outcomes in older individuals. Some previous research has demonstrated that poor glycemic control and longer duration of T2DM are associated with cognitive impairment. However, our study did not find statistically significant associations between the related indicators of diabetes and recognize impairment. Reviewing this study, we suspect that this may be attributed to potential selection bias in subjects. All the enrolled subjects were inpatients with a critically ill state and high Hba1c levels.

Our study revealed a significant positive association between educational level and cognitive function, consistent with previous research findings[25]. Additionally, our study indicated a marginal protective impact of moderate alcohol on cognitive function. Some evidence, though inconclusive, primarily from observational studies and sporadically from clinical trials, suggests a protective association between moderate alcohol consumption and cognitive outcomes in older individuals[26; 27]. Previous research has demonstrated that poor glycemic control and longer duration of T2DM are associated with cognitive impairment. [28; 29] However, our study did not find statistically significant associations between the diabetes-related indicators and cognitive impairment. We speculate that potential selection bias in subjects may account for this, as all participants in our study were hospitalized, presenting a poor condition and high HbA1c levels.

This study also had some limitations. Firstly, the inadequate sample size might affect the generalizability of the model. Secondly, including only hospitalized T2DM patients may introduce potential study bias. Further research is needed to investigate the progression from MCI to AD in individuals with T2DM.

Radiomics, based on the XGBoost algorithm, could serve as an auxiliary tool for identifying early cognitive impairment in type 2 diabetes mellitus.

T2DM: Type 2 diabetes mellitus

MoCA: Montreal Cognitive Assessment

MCI: Mild cognitive impairment

AUC: Area under the curve

ROC Receiver operating characteristic

XGBoost: eXtreme Gradient Boosting Machine

SHAP: SHapley Additive exPlanations

Funding

The work was supported by the National Natural Science Foundation of China (Grant No.82101983) and the Medical Health Science and Technology Project of Zhejiang Province (2021KY472).

Acknowledgements

None

Competing Interests

The authors declare no conflicts of interest related to the publication of this article.

Ethics approval

The research obtained approval from the Ethics Committee of Zhejiang Provincial People's Hospital.

Consent to participate

Written informed consent was obtained from all participants.

Author Contributions

Yuan Shao contributed to Conceptualization, Methodology, Formal analysis, Writing - Original Draft, and Writing - Review & Editing. Chaofei Gu contributed to Conceptualization, Methodology, Data curation, and Writing - Review & Editing. Hanwen Xu contributed to Formal analysis, Data curation, and Writing - Review & Editing. Sb contributed to Methodology and Formal analysis. Zhenyu Shu contributed to Supervision and Writing. Yingxiang Song contributed to Conceptualization, Methodology, Data curation, Formal analysis, Supervision, and Writing - Review & Editing.

Data availability statement

All data generated or analysed during this study are included in this published article [and its supplementary information files].

A. Moheet, S. Mangia, and E.R. Seaquist, Impact of diabetes on cognitive function and brain structure. Annals of the New York Academy of Sciences 1353 (2015) 60–71.
F.P. Lu, K.P. Lin, and H.K. Kuo, Diabetes and the risk of multi-system aging phenotypes: a systematic review and meta-analysis. PloS one 4 (2009) e4144.
A.A. Willette, B.B. Bendlin, E.J. Starks, A.C. Birdsill, S.C. Johnson, B.T. Christian, O.C. Okonkwo, A. La Rue, B.P. Hermann, R.L. Koscik, E.M. Jonaitis, M.A. Sager, and S. Asthana, Association of Insulin Resistance With Cerebral Glucose Uptake in Late Middle-Aged Adults at Risk for Alzheimer Disease. JAMA neurology 72 (2015) 1013–20.
J.J. Manly, M.X. Tang, N. Schupf, Y. Stern, J.P. Vonsattel, and R. Mayeux, Frequency and course of mild cognitive impairment in a multiethnic community. Annals of neurology 63 (2008) 494–506.
C.R. Jack, Jr., D.S. Knopman, W.J. Jagust, R.C. Petersen, M.W. Weiner, P.S. Aisen, L.M. Shaw, P. Vemuri, H.J. Wiste, S.D. Weigand, T.G. Lesnick, V.S. Pankratz, M.C. Donohue, and J.Q. Trojanowski, Tracking pathophysiological processes in Alzheimer's disease: an updated hypothetical model of dynamic biomarkers. The Lancet. Neurology 12 (2013) 207–16.
C. Sims-Robinson, B. Kim, A. Rosko, and E.L. Feldman, How does diabetes accelerate Alzheimer disease pathology? Nature reviews. Neurology 6 (2010) 551–9.
K. Blennow, and H. Hampel, CSF markers for incipient Alzheimer's disease. The Lancet. Neurology 2 (2003) 605–13.
X. Zhao, Q. Han, Y. Lv, L. Sun, X. Gang, and G. Wang, Biomarkers for cognitive decline in patients with diabetes mellitus: evidence from clinical studies. Oncotarget 9 (2018) 7710–7726.
G. Lombardi, G. Crescioli, E. Cavedo, E. Lucenteforte, G. Casazza, A.G. Bellatorre, C. Lista, G. Costantino, G. Frisoni, G. Virgili, and G. Filippini, Structural magnetic resonance imaging for the early diagnosis of dementia due to Alzheimer's disease in people with mild cognitive impairment. The Cochrane database of systematic reviews 3 (2020) Cd009628.
Y.D. Reijmer, M. Brundel, J. de Bresser, L.J. Kappelle, A. Leemans, and G.J. Biessels, Microstructural white matter abnormalities and cognitive functioning in type 2 diabetes: a diffusion tensor imaging study. Diabetes care 36 (2013) 137–44.
M. Benzagmout, S. Boujraf, B. Alami, H.A. Amadou, H. El Hamdaoui, A. Bennani, M. Jaafari, I. Rammouz, M. Maaroufi, R. Magoul, and D. Boussaoud, Emotion processing in Parkinson's disease: a blood oxygenation level-dependent functional magnetic resonance imaging study. Neural regeneration research 14 (2019) 666–672.
F.J. Meijer, and B. Goraj, Brain MRI in Parkinson's disease. Frontiers in bioscience (Elite edition) 6 (2014) 360–9.
S.Y. Won, Y.W. Park, M. Park, S.S. Ahn, J. Kim, and S.K. Lee, Quality Reporting of Radiomics Analysis in Mild Cognitive Impairment and Alzheimer's Disease: A Roadmap for Moving Forward. Korean journal of radiology 21 (2020) 1345–1354.
S. Shinde, S. Prasad, Y. Saboo, R. Kaushick, J. Saini, P.K. Pal, and M. Ingalhalikar, Predictive markers for Parkinson's disease using deep neural nets on neuromelanin sensitive MRI. NeuroImage. Clinical 22 (2019) 101748.
R. Shahidi, M. Baradaran, A. Asgarzadeh, S. Bagherieh, Z. Tajabadi, A. Farhadi, S.S. Korani, M. Khalafi, P. Shobeiri, H. Sadeghsalehi, A. Shafieioun, M.A. Yazdanifar, A. Singhal, and H. Sotoudeh, Diagnostic performance of MRI radiomics for classification of Alzheimer's disease, mild cognitive impairment, and normal subjects: a systematic review and meta-analysis. Aging clinical and experimental research 35 (2023) 2333–2348.
Z.S. Nasreddine, N.A. Phillips, V. Bédirian, S. Charbonneau, V. Whitehead, I. Collin, J.L. Cummings, and H. Chertkow, The Montreal Cognitive Assessment, MoCA: a brief screening tool for mild cognitive impairment. Journal of the American Geriatrics Society 53 (2005) 695–9.
V. Hachinski, S. Oveisgharan, A.K. Romney, and W.R. Shankle, Optimizing the Hachinski Ischemic Scale. Archives of neurology 69 (2012) 169–75.
C.P. Hughes, L. Berg, W.L. Danziger, L.A. Coben, and R.L. Martin, A new clinical scale for the staging of dementia. The British journal of psychiatry: the journal of mental science 140 (1982) 566–72.
M.P. Lawton, and E.M. Brody, Assessment of older people: self-maintaining and instrumental activities of daily living. The Gerontologist 9 (1969) 179–86.
D.B. Maurel, N. Boisseau, C.L. Benhamou, and C. Jaffre, Alcohol and bone: review of dose effects and mechanisms. Osteoporosis international: a journal established as result of cooperation between the European Foundation for Osteoporosis and the National Osteoporosis Foundation of the USA 23 (2012) 1–16.
A.M. Wood, S. Kaptoge, A.S. Butterworth, P. Willeit, S. Warnakula, T. Bolton, E. Paige, D.S. Paul, M. Sweeting, S. Burgess, S. Bell, W. Astle, D. Stevens, A. Koulman, R.M. Selmer, W.M.M. Verschuren, S. Sato, I. Njølstad, M. Woodward, V. Salomaa, B.G. Nordestgaard, B.B. Yeap, A. Fletcher, O. Melander, L.H. Kuller, B. Balkau, M. Marmot, W. Koenig, E. Casiglia, C. Cooper, V. Arndt, O.H. Franco, P. Wennberg, J. Gallacher, A.G. de la Cámara, H. Völzke, C.C. Dahm, C.E. Dale, M.M. Bergmann, C.J. Crespo, Y.T. van der Schouw, R. Kaaks, L.A. Simons, P. Lagiou, J.D. Schoufour, J.M.A. Boer, T.J. Key, B. Rodriguez, C. Moreno-Iribas, K.W. Davidson, J.O. Taylor, C. Sacerdote, R.B. Wallace, J.R. Quiros, R. Tumino, D.G. Blazer, 2nd, A. Linneberg, M. Daimon, S. Panico, B. Howard, G. Skeie, T. Strandberg, E. Weiderpass, P.J. Nietert, B.M. Psaty, D. Kromhout, E. Salamanca-Fernandez, S. Kiechl, H.M. Krumholz, S. Grioni, D. Palli, J.M. Huerta, J. Price, J. Sundström, L. Arriola, H. Arima, R.C. Travis, D.B. Panagiotakos, A. Karakatsani, A. Trichopoulou, T. Kühn, D.E. Grobbee, E. Barrett-Connor, N. van Schoor, H. Boeing, K. Overvad, J. Kauhanen, N. Wareham, C. Langenberg, N. Forouhi, M. Wennberg, J.P. Després, M. Cushman, J.A. Cooper, C.J. Rodriguez, M. Sakurai, J.E. Shaw, M. Knuiman, T. Voortman, C. Meisinger, et al., Risk thresholds for alcohol consumption: combined analysis of individual-participant data for 599 912 current drinkers in 83 prospective studies. Lancet (London, England) 391 (2018) 1513–1523.
G. Dissaux, D. Visvikis, R. Da-Ano, O. Pradier, E. Chajon, I. Barillot, L. Duvergé, I. Masson, R. Abgral, M.J. Santiago Ribeiro, A. Devillers, A. Pallardy, V. Fleury, M.A. Mahé, R. De Crevoisier, M. Hatt, and U. Schick, Pretreatment (18)F-FDG PET/CT Radiomics Predict Local Recurrence in Patients Treated with Stereotactic Body Radiotherapy for Early-Stage Non-Small Cell Lung Cancer: A Multicentric Study. Journal of nuclear medicine: official publication, Society of Nuclear Medicine 61 (2020) 814–820.
K.C. Yuan, L.W. Tsai, K.H. Lee, Y.W. Cheng, S.C. Hsu, Y.S. Lo, and R.J. Chen, The development an artificial intelligence algorithm for early sepsis diagnosis in the intensive care unit. International journal of medical informatics 141 (2020) 104176.
F. Yi, H. Yang, D. Chen, Y. Qin, H. Han, J. Cui, W. Bai, Y. Ma, R. Zhang, and H. Yu, XGBoost-SHAP-based interpretable diagnostic framework for alzheimer's disease. BMC medical informatics and decision making 23 (2023) 137.
M. Lövdén, L. Fratiglioni, M.M. Glymour, U. Lindenberger, and E.M. Tucker-Drob, Education and Cognitive Functioning Across the Life Span. Psychological science in the public interest: a journal of the American Psychological Society 21 (2020) 6–41.
N. Scarmeas, C.A. Anastasiou, and M. Yannakoulia, Nutrition and prevention of cognitive impairment. The Lancet. Neurology 17 (2018) 1006–1015.
M. Koch, A.L. Fitzpatrick, S.R. Rapp, R.L. Nahin, J.D. Williamson, O.L. Lopez, S.T. DeKosky, L.H. Kuller, R.H. Mackey, K.J. Mukamal, M.K. Jensen, and K.M. Sink, Alcohol Consumption and Risk of Dementia and Cognitive Decline Among Older Adults With or Without Mild Cognitive Impairment. JAMA network open 2 (2019) e1910319.
S.A. Ebady, M.A. Arami, and M.H. Shafigh, Investigation on the relationship between diabetes mellitus type 2 and cognitive impairment. Diabetes research and clinical practice 82 (2008) 305–9.
S.M. Manschot, A.M. Brands, J. van der Grond, R.P. Kessels, A. Algra, L.J. Kappelle, and G.J. Biessels, Brain magnetic resonance imaging correlates of impaired cognition in patients with type 2 diabetes. Diabetes 55 (2006) 1106–13.

Table 1 to 5 are available in the Supplementary Files section.

No competing interests reported.

Download PDF

Reviewers invited by journal
18 Jun, 2024
Editor assigned by journal
18 Jun, 2024
Editor invited by journal
11 Apr, 2024
Submission checks completed at journal
10 Apr, 2024
First submitted to journal
24 Mar, 2024

You are reading this latest preprint version

XGBoost-SHAP-based interpretable diagnostic framework for early cognitive impairment in type 2 diabetes mellitus

Status:

Version 1

Abstract

Figures

Key points

Introduction

Methods

Study design and participants

Acquisition and analysis of MR images

Acquisition and selection of radiomics features

Radiomics model construction and validation

Statistical Analysis

Ethics approval

Results

Patient characteristics

Performance of a clinical model

Radiomic feature selection and performance of radiomics model

Development and visualization of radiomics-clinical combined models

Discussion

Conclusions

Abbreviations

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1