This study comprised two primary analyses: a genetic-centric analysis (Analysis 1; detailed in Fig.1 and the Methods section) and a genetic-imaging integrative analysis (Analysis 2; detailed in Fig.2 and the Methods section). Data used in the two analyses are summarized (Supplementary TableS1). A total of 68,911 participants from the TWB were included in the analysis (Fig. S1).

The dataset containing information from 60,747 individuals after data quality control (QC) was divided into several subsets: (i) The genome-wide association study (GWAS) samples (Dataset 1, N=35,688), training samples (Dataset 2, N=12,236; Dataset 4, N=40,787), and validation samples (Dataset 3, N=3060; Dataset 5, N=10,197). For classification analysis, testing samples comprised Dataset 6 (N=8827) and Dataset 7 (N=936), while for prediction analysis, they were represented as Datasets 6 (N=8827) and Dataset 7 (N=936); B Sample size. Total sample size, along with the number of cases and the number of controls, are shown for each of the four phenotype definitions in Datasets 1 7; C Phenotype definition criteria. The definition and sample size for the four Type 2 Diabetes (T2D) phenotype definitions is shown. D Analysis flowchart. The analysis flow comprises three steps, starting with selecting T2D-associated single nucleotide polymorphisms (SNPs) and polygenic risk score (PRS), then selecting demographic and environmental covariates, and the best XGBoost model was established using the selected features. As to the first step, SNPs can be chosen from A our own GWAS with an adjustment for age, sex, and top ten principal components (PCs), B published studies based on single ethnic populations, and C published studies based on multiple ethnic populations. Source data are provided as a Source Data file.

Phenotype Definition IV was used as an example to illustrate the process. The data containing information from 7,786 individuals were divided into four subsets: a training dataset (N=4689), a validation dataset (N=1175), and two independent testing datasets (N=1469 for the first dataset and N=444 for the second independent dataset). Subsequently, the best XGBoost model was established. B Flowchart of PRS construction. The Polygenic Risk Score (PRS) was constructed using PRS-CSx, utilizing genome-wide association study (GWAS) summary statistics from the European (EUR), East Asian (EAS), and South Asian (SAS) populations obtained from the analysis of the DIAGRAM Project. Source data are provided as a Source Data file.

We evaluated the prediction performance under different scenarios hierarchically (the best scenario at a previous variable was given for a discussion of the next variable) in the following order: the sources and significance levels of T2D-associated SNPs (Fig.3A and Fig. S2), T2D phenotype definitions (Fig.3B), family history variable combinations (Fig.3C and Fig. S3), demographic variable combinations (Fig.3D), demographic and genetic variable combinations (Fig.3E), and SNP and PRS combinations (Figs.3F and 3G). The findings are summarized as follows: First, using T2D-associated SNPs from the previous large-sample-size GWAS11 as predictors had the highest AUC of 0.557, but its AUC was not significantly higher than that used the SNPs identified by our smaller-sample-size GWAS under different thresholds of statistical significance (Fig.3A), although our GWASs did identify some T2D-associated SNPs (Fig. S4). Second, the phenotype defined by self-reported T2D with HbA1C6.5% or fasting glucose 126mg/dL (i.e., T2D Definition IV) had the highest AUC of 0.640. Its AUC was significantly higher than the AUCs of the other three T2D definitions (Fig.3B). Third, sibs disease history had a significantly higher AUC of 0.732 than parents disease history with an AUC of 0.670 (p=0.009). Moreover, additive parent-and-sib disease history had the highest AUC of 0.758. Its AUC was significantly higher than parent-only (p<0.001) (Fig.3C). Fourth, a joint effect of age, sex, and additive parent-sib disease history had the highest AUC of 0.884. Its AUC was significantly higher than other demographic variable combinations, except for the combination of age and additive parent-sib disease history (Fig.3D). Fifth, whatever SNPs were included or not, demographic and PRS combinations outperformed the models without incorporation of PRS (Fig.3E), although genetic factors only improved up to 3% of AUC conditional on demographic characteristics (age, sex, and family history of T2D). Finally, given T2D-associated SNPs, AUC significantly increased if PRS was included (Fig.3F); T2D-associated SNPs provided a limited additional effect if PRS was already included (Fig.3G).

A bar chart displays AUC. The two-sided DeLong test examined the difference between AUCs. Bonferronis correction was applied to control for a family-wise error rate in multiple comparisons. Symbols *, **, and *** indicate p-values<0.05, 0.01, and 0.001, respectively. A SNP selection. Model predictors were SNPs selected from published studies or our GWAS under different p-value thresholds, where our GWAS association test is a two-sided Wald test for the slope coefficient in a logistic regression. The average AUCs of prediction models for four phenotype definitions were compared. B T2D Phenotype Definition. In addition to including the selected variables in Fig.3A, the AUCs of four phenotype definitions were compared. C Family history of T2D. In addition to including the selected variables in Fig.3A, B, the AUCs of the four types of T2D family history (i.e., (i): parents (binary factor), (ii) sibs (binary factor), (iii) either parents or sibs (binary factors), and (iv) both parents and sibs (ordinal factor)) were compared. D Demographic variables. In addition to including the selected variables in Fig.3AC, the AUCs of different combinations of demographic factors, including age, sex, and family history of T2D, are compared. E PRS and demographic variables. In addition to including the selected variables in Fig.3AD, the AUCs of different combinations of genetic variables, including SNPs, PRS-CS, and PRS-CSx, and demographic variables, including age, sex, and family history of T2D, are compared. F Impact of including PRS after SNPs. The AUCs of the models that consider SNPs, SNPs+PRS-CS, and SNPs+PRS-CSx as predictors are compared. G Impact of including additional SNPs after PRS. The additional 137 SNPs were collected from published studies (Supplemental Text2). The AUCs of the models that consider additional SNPs given PRS in the model are compared. Source data are provided as a Source Data file.

Among different prediction models, the model with predictors PRS-CSx, age, sex, and family history of T2D had the highest AUC 0.915 (Fig.4A) for Type VI definition of T2D based on the first testing dataset (i.e., Dataset 6 in Fig.1). The optimal threshold, determined by the Youden index, for the fitted value that used to predict T2D or non-T2D in the XGboost model was 0.16. The Accuracy, Sensitivity, Specificity, and F1 indices were 0.843, 0.844, 0.843, and 0.672, respectively. Furthermore, the model was tested in the second independent testing dataset (i.e., Dataset 7 in Fig.1), and a promising result similar to the first testing dataset was found: AUC=0.905, Accuracy = 0.843, Sensitivity = 0.846, Specificity = 0.842, and F1=0.644. AUCs are also provided for the other three T2D definitions (Fig. S5).

A AUCs of all models based on Phenotype Definition IV. A heatmap summarizes the AUCs of all models based on Phenotype Definition IV (i.e., T2D was defined by self-reported T2D, HbA1c, and fasting glucose). The genetic variables are shown on the X-axis, and the demographic variables are shown on the Y-axis. B Positive correlation between PRS and T2D odds ratio. In each decile of PRS based on PRS-CSx, the odds ratio of T2D risk and its 95% confidence interval were calculated based on an unadjusted model (blue line) and an adjusted model considering age, sex, and T2D family history (red line). The reference group was the PRS group in the 4060% decile. The horizontal bars are presented as the odds ratio estimates (square symbol) +/ its 95% confidence intervals (left and right ends) at a PRS decile. C High-risk group. In the chart, the figures from the inner to the outer represent (i) the case-to-control ratio, (ii) the number of cases, and (iii) the number of controls in the PRS decile subgroups. D Association of age, sex, T2D family history, and PRS with T2D. In the univariate analysis, the p-values for age, sex, family history, and PRS were 4.17 1020, 7.08 107, 9.41 1013, and 2.06 1013, respectively. In the multivariate analysis, the p-values for age, sex, family history, and PRS were 2.00 1016, 5.56 105, 1.43 1010, and 5.49 1013, respectively. E Risk factors for T2D. Kaplan-Meier curves reveal that Age (older individuals), sex (males), T2D family history (the larger number of parents and siblings who had T2D), and PRS (high decile PRS group) are risk factors (high-risk level) for T2D risk. F Median event time of T2D. Examples of the median event time for developing T2D are provided based on a multivariate Cox regression model, both without and with incorporating PRS. NA indicates not assessable. Source data are provided as a Source Data file.

The importance of each predictor was evaluated through a backward elimination procedure of variables. The optimal model incorporating age, sex, family history of T2D, and PRS achieved an AUC of 0.915. The AUC reductions upon removing individual variables are as follows: (a) Omitting the age variable resulted in an AUC of 0.839, representing a reduction of 0.076. (b) Excluding the sex variable resulted in an AUC of 0.905, with a decrease of 0.01. (c) Removing the family history of the T2D variable yielded an AUC of 0.881, with a reduction of 0.034. (d) Eliminating the PRS variable resulted in an AUC of 0.884, decreasing to 0.031. Based on the decrease in AUC, the impact size appears to be in the order of age > family history > PRS > sex. Additionally, we evaluated feature importance (see the Methods section), and the order of feature importance is family history > age > PRS > sex. Our findings consistently highlight age and family history as the most crucial risk factors for T2D.

Family history encompasses genetics and environment. We delved into the connection between the family history of T2D treated as a graded scale (0, 1, 2, 3, and 4) and the genetic component represented by the PRS. Through ordinal logistic regression, we observed a beta coefficient of 0.808 and an associated odds ratio (OR) of 2.24 (p=1.65 10296). The remarkably small p-value emphasizes the robust statistical significance, signaling a substantial association between the PRS and familial T2D status. For each incremental unit rise in an individuals PRS, their odds of belonging to a higher family history category for T2D increase by 2.24 times. This implies a tangible shift in the likelihood of different family history classifications as the PRS changes. The findings underscore a strong statistical link between genetic predisposition, as captured by the PRS, and the gradation of family history of T2D.

Furthermore, we calculated the Population Attributable Risk (PAR) by dichotomizing PRS into a high-risk group (PRS tercile >80%) and a non-high-risk group (PRS tercile <80%). Among the 59,811 participants, the breakdown was as follows: high PRS with family history (N=5473), high PRS without family history (N=6489), non-high PRS with family history (N=16,054), and non-high PRS without family history (N=31,795). The PAR estimate was 10.17%, indicating that 10.17% of the family history of T2D is attributed to genetic heritability. If considering a broader definition of the high-risk group (PRS tercile >60%) and non-high-risk group (PRS tercile <60%), the PAR estimate increased to 18.41%.

Further consideration of environmental factors, including education level, drinking experience, exercise habits, the number of exercise types, and SNP-SNP interactions with and without SNPs main effect, did not improve T2D prediction (Supplementary TableS2). Considering model parsimony, the final model did not include these environmental factors and SNP-SNP interactions. In addition to prediction models, classification models were also established. The AUCs in classification models (Fig. S6) were generally slightly higher than those in prediction models (Fig. S5).

The positive association between PRS and T2D risk is shown (Fig.4B). Compared to the participants in the 4060% PRS decile group, those in the top 10% decile group had a 4.738-fold risk of developing T2D (95% confidence interval: 3.1477.132, p<0.001) and a 4.660-fold risk (95% confidence interval: 2.6828.097, p<0.001) after adjusting for age, sex, and family history. In addition, we performed a stratified analysis across various combinations of age subgroups, sex subgroups, and family history subgroups to identify high-risk subgroups, where age was stratified into four subgroups based on quartiles: 025%, 2550%, 5075%, and 75100%, corresponding to age subgroups of 43, 4352, 5259, and >59 years of age, respectively (Fig. S7). We identified a high-risk subgroup of women who were older than 59 and had a family history of T2D. The ratio of case vs. control sample size was as high as 7.313.0-fold in the 80100% decile groups (Fig.4C). The ratio was much higher than a 1.6-fold that did not consider PRS (i.e., PRS at 0100%) (Fig.4C). Due to ambiguity or instability in the evidence for other combinations, we chose not to report them.

Among 8347 non-T2D participants at baseline in the first testing dataset of 8827 participants, 220 reported T2D in the follow-up. The Cox regression analyses considered two types of time scales and three types of sex variable treatment and obtained a consistent result (Supplementary TableS3). Using the analysis in which we considered time-on-study as the time-scale with age at baseline, sex, family history of T2D, and PRS as covariates for illustration, age, sex, family history of T2D, and PRS were all significantly associated with T2D (p<0.001) (Fig.4D). Increased age, higher PRS, and stronger T2D family history had a higher T2D risk. The elderly male, with a strong family history and high PRS, had a severe T2D risk (Fig.4E for multivariate Cox regression and Fig. S8 for univariate Cox regression). We also provided the predicted time-to-event (week) (Fig.4F). For example, a 50-year-old man with one of his family members had T2D will achieve median T2D-free time after 460 weeks (95% CI, 384NA). The median time to develop T2D was shortened to 419 weeks (95% CI, 384NA) after considering a standardized PRS of 0.66 (equivalent to a PRS risk subgroup in the top 25% of the population).

A linear regression analysis was performed to assess the impact of exercise on HbA1c. Multiple testing for 110 analyses was corrected using Bonferroni correction, and the significance level was set as 4.5 10-4. It was observed that individuals engaging in regular exercise experienced a significant reduction in HbA1c by an average of 0.09% mg/dL (p<0.001) compared to those who did not engage in regular exercise. Moreover, individuals with a high PRS who engaged in exercise demonstrated a greater reduction in HbA1c (0.13% mg/dL) than those with a low PRS (0.08% mg/dL). The results also suggested that the T2D patients who regularly engaged in exercise can have a noteworthy improvement of 0.32% mg/dL in HbA1c than those T2D patients who did not exercise regularly. In addition, among the various types of exercise, walking for fitness exhibited the most robust reduction in HbA1c for all samples, including high and low-risk subgroups and both T2D and non-T2D groups (Fig. S9). On average, participants engaged in walking for fitness 18.30 times a month (standard deviation = 8.64) for approximately 48.13minutes per session (standard deviation = 22.92).

To investigate the early detection capability of our model for T2D, we performed an analysis focusing on 550 women participants older than 59 years, all of whom had a family history of T2D. We identified them as at high risk if they possessed a high PRS, even though they were initially reported as non-T2D at baseline. Thirty-six were changed to T2D, and 514 were still non-T2D at follow-up. We predicted their T2D status. G1 G4 are the groups of participants in true positive, false negative, false positive, and true negative, respectively (Fig.5A). We evaluated that G3 was indeed misclassified by our prediction model or our prediction had corrected the problem in the self-reported T2D by further investigating: (1) their follow-up time and current risk in the Cox regression model; (2) HbA1c and fasting glucose; (3) the accuracy of self-reported disease status.

A Four subgroups (N=550). B Survival rate (N=550). C Median survival time (N=550). P-values of G1 vs. G3, G2 vs. G3, and G4 vs. G3 were 0.092, 0.0014 (**), and 2.22 1016 (***), respectively. D Follow-up time (N=550). P-values of G1 vs. G3, G2 vs. G3, and G4 vs. G3 were 0.056, 0.32, and 0.14, respectively. E T2D risk (N=550). P-values of G1 vs. G3, G2 vs. G3, and G4 vs. G3 were 0.018 (*), 0.073, and 0.0039 (**), respectively. F HbA1c (N=550). P-values of G1 vs. G3, G2 vs. G3, and G4 vs. G3 were 2.21 1014, 0.0039, and 3.00 105; respectively; in the follow-up, p-values of G1 vs. G3, G2 vs. G3, and G4 vs. G3 were 1.50 10-13, 6.01 104, and 4.55 10-6, respectively. G Fasting glucose (N=550). In the baseline, p-values of G1 vs. G3, G2 vs. G3, and G4 vs. G3 were 2.06 10-12, 6.66 104, and 1.63 102; respectively; in the follow-up, p-values of G1 vs. G3, G2 vs. G3, and G4 vs. G3 were 8.30 108, 1.38 103, and 1.84 102, respectively. H Phenotype definition in G3 (N=395). Many individuals in G3 cannot satisfy the T2D Phenotype Definition IV. In Fig.5CG, two-sided Wilcoxon rank-sum tests were applied to compare group differences. The box plots center lines indicate the medians, the lower and upper boundaries of the boxes represent the first and third quartiles, and the whiskers extend to cover a range of 1.5 interquartile distances from the edges. The violin plots upper and lower bounds depict the minimum and maximum values. Source data are provided as a Source Data file.

The Kaplan-Meier curve for each subgroup is depicted (Fig.5B). The distributions of median survival time for each subgroup are illustrated (Fig.5C). The distributions of the time period from baseline to follow-up for each subgroup are presented (Fig.5D). The distributions of Type 2 diabetes (T2D) risk at follow-up for each subgroup are shown (Fig.5E). The distributions of HbA1c levels at baseline and follow-up for each subgroup are displayed (Fig.5F). The distributions of fasting glucose levels at baseline and follow-up are demonstrated (Fig.5G).

First, compared to G4 (true negative), G3 had a significantly lower T2D-free probability (Fig.5B), shorter median survival time (Fig.5C), higher T2D-risk under similar follow-up time (Fig.5D and 5E), higher HbA1c (Fig.5F), and higher fasting glucose (Fig.5G). Second, compared to G1 (true positive), G3 had a comparable survival rate (Fig.5B), median survival time (Fig.5C), and T2D-risk under similar follow-up time (Fig.5D and5E) but lower HbA1c (Fig.5F) and fasting glucose (Fig.5G). We didnt compare G2 and G3 because of the small sample size in G2. Finally, among the 395 participants in G3, 80.76% of them were removed from our previous analysis because their baseline HbA1c and fasting glucose violated the criteria for the phenotype definition (Fig.1C); 339 participants were removed because of their follow-up HbA1c and fasting glucose violated the formal T2D criteria; only 34 self-reported non-T2D were really non-T2D participants who had HbA1C<6.5% and fasting glucose <126mg/dL (Fig.5H). Overall, the results consistently indicate that G3 represents individuals in a pre-T2D stage, which our model can detect early.

The model that combined four types of image features performed best. Moreover, the model based on BMD image features exhibited a higher AUC, accuracy, specificity, and F1 than the models based on any other three types of images (Fig.6A). The models based on image features had an AUC of 0.898, higher than the ones of genetic information (AUC=0.677) and demographic factors (AUC=0.843). Integrating image features, genetic information, and demographic factors increased AUC to 0.949 in the first testing data (Fig.6B); the results for each of the four images are also provided (Fig. S10). The accuracy, sensitivity, specificity, and F1 of the model in the first testing data were 0.871, 0.878, 0.870, and 0.663, respectively, based on a classification threshold of 0.03. The model also performed reasonably well in the second testing dataset with AUC=0.929, Accuracy = 0.854, Sensitivity = 0.789, Specificity = 0.862, and F1=0.558. The results of a prediction model using tuned parameters are also provided (Supplementary TableS4). As no significant improvement was observed, this paper discusses the default model. According to the estimated feature importance in the best XGBoost model, all genetic factors (PRS), four types of medical images, and demographic variables provided informative features for risk assessment, such as PRS (genetics), family history and age (demographic factors), fatty liver (ABD images), end-diastolic velocity in the right common carotid artery (CAU images), RR interval (ECG images), and spine thickness (BMD images). Of the 152 medical imaging features, 125 were selected in the final model. (Fig.6C).

A Performance comparison of medical imaging data analysis. The area under the receiver operating characteristic (ROC) curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPEC), and F1 score are compared for the integrative analysis of four types of medical images (All) and individual medical image analyses, including BMD, ECG, CAU, and ABD. B The model that combines four types of medical imaging, PRS, and demographic variables shows the highest AUC of 0.949. ROC plots and the corresponding AUC for the models considering medical image features (I), genetic PRS (G), and demographic variables, including age, sex, T2D family history (D), and their combinations. C An optimal model combining medical imaging, PRS, and demographic variables. The best models top 20 features with a high feature impact include the medical image, genetic, and demographic features. D Positive correlation between MRS and T2D odds ratio. In each decile of MRS based on four types of medical images, the odds ratio of T2D risk and its 95% confidence interval were calculated based on an unadjusted model (blue line) and an adjusted model considering age, sex, and T2D family history (red line), with the MRS group in the 4060% decile serving as the reference group. The horizontal bars are presented as the odds ratio estimates (square symbol) +/ its 95% confidence intervals (left and right ends). E High-risk group. The figures from the inner to the outer in the chart display (i) the case-to-control ratio, (ii) the number of cases, and (iii) the number of controls in the MRS decile subgroups. F Input page of the online T2D prediction website. Personal information, including age, sex, family history of T2D, PRS, and MRS, is input to calculate T2D risk. PRS and MRS are optional, and a reference distribution is provided. G Output page of the online T2D prediction website. Source data are provided as a Source Data file.

To address the challenges of practical clinical implementation in the best XGBoost model, we have proposed an alternative model that requires a limited number of features. We systematically calculated each features incremental area under the AUC by sequentially including those with the highest feature importance. We selected the top features showing a positive AUC increment. The analysis revealed that a sub-model incorporating only the following eight crucial variables: family history (from the questionnaire), age (from the questionnaire), fatty liver (from ABD images), spine thickness (from BMD images), PRS (from genetic data), end-diastolic velocity in the right common carotid artery (R_CCA_EDV) (from CAU images), RR interval (from ECG images), and end-diastolic velocity in the left common carotid artery (L_CCA_EDV) (from CAU images), maintains a commendable AUC of 0.939 (Fig. S11). This streamlined model significantly reduces the number of risk predictors while preserving high prediction accuracy, demonstrating promising potential for practical application in clinical settings. Moreover, the reduced number of risk predictors in the streamlined model alleviates concerns about model overfitting.

Each participants multi-image risk score (MRS) was calculated as the likelihood of being predicted as a T2D case using XGBoost, which analyzed the medical imaging features for T2D prediction. The odds ratio and its confidence interval for the association between MRS and T2D are shown by percentiles of MRS (Fig.6D). Compared to the participants in the 4060% MRS decile group, the risk of T2D increased with MRS. Of importance, we further identified that, for the men older than 54 years old with a family history of T2D, the case vs. control ratio of sample size was 9.3 in the 90100% MRS decile group, much higher than 1.3, which MRS was not considered (Fig.6E).

We have established a website where users can calculate their T2D risk online. To obtain the risk assessment, users are required to provide age, sex, and family history of T2D, and they can optionally provide PRS and MRS (Fig.6F). PRS and MRS can be entered manually or uploaded as a file (Supplemental Text1). Additionally, we have provided PRS and MRS risk percentages based on the study population as a reference. The online risk assessment offers information, including the risk of developing T2D over 3, 5, and 7 years, T2D-free probability, and T2D risk with and without considering PRS (Fig.6G). The assessment takes into account both PRS and MRS (Fig.6G). For example, consider a 50-year-old male with a family history of T2D and PRS 1.5 and MRS 1.5. Without considering PRS, the risk (probability) of developing T2D after a 7-year follow-up is 0.23. However, when PRS is considered, the risk increases to 0.37. Furthermore, considering MRS further increases the risk to 0.81. The online tool provides these valuable insights to users based on their input data.

Link:
AI-enhanced integration of genetic and medical imaging data for risk assessment of Type 2 diabetes - Nature.com