Visceral Adiposity and Progression of ADPKD: A Cohort Study of Patients from the TEMPO 3:4 Trial

Abstract

Rationale & Objective:

Body-mass index (BMI) is an independent predictor of kidney disease progression in individuals with autosomal dominant polycystic kidney disease (ADPKD). Adipocytes do not simply act as a fat reservoir but are active endocrine organs. We hypothesized that greater visceral abdominal adiposity would associate with more rapid kidney growth in ADPKD and influence the efficacy of tolvaptan.

Study Design:

A retrospective cohort study.

Setting & Participants:

1053 patients enrolled in the TEMPO 3:4 tolvaptan trial with ADPKD and high risk of rapid disease progression.

Predictor:

Estimates of visceral adiposity extracted from coronal plane MRIs using deep learning.

Outcome:

Annual change in total kidney volume (TKV) and effect of tolvaptan on kidney growth.

Analytical Approach:

Multinomial logistic regression and linear mixed models.

Results:

In fully adjusted models, the highest tertile of visceral adiposity was associated with greater odds of annual change in TKV of ≥7% vs. <5% (OR: 4.78 [3.03, 7.47]). The association was stronger in females than males (interaction p<0.01). In linear mixed models with an outcome of % change in TKV per year, tolvaptan efficacy (% change in TKV) was reduced with higher visceral adiposity (three-way interaction of treatment*time*visceral adiposity p=0.002). Visceral adiposity significantly improved classification performance of predicting rapid annual % change in TKV for individuals with a normal BMI (De-Long’s test Z-score: −2.03; p=0.04). Greater visceral adiposity was not associated with estimated glomerular filtration rate (eGFR) slope in the overall cohort; however, visceral adiposity was associated with more rapid decline in eGFR slope (below the median) in females (fully adjusted OR 1.06 [1.01, 1.11] per 10 unit increase in visceral adiposity) but not males (0.98 [0.95, 1.02]).

Limitations:

Retrospective; rapid progressors; computational demand of deep learning.

Conclusions:

Visceral adiposity that can be quantified by MRI in the coronal plane using a deep learning segmentation model, independently associates with more rapid kidney growth, and improves classification of rapid progression in individuals with a normal BMI. Tolvaptan efficacy decreases with increasing visceral adiposity.

Plain Language Summary

Visceral Adiposity Predicts Kidney Growth in Patients with ADPKD.

We analyzed images from a previous study with the drug, tolvaptan, conducted in patients with autosomal dominant polycystic kidney disease (ADPKD), to measure the amount of fat tissue surrounding the kidneys (visceral fat). Because we had previously shown body-mass index can predict kidney growth in this population, we determined whether visceral fat was an important factor associated with kidney growth. Using a machine learning tool to automate measurement of fat in images, we observed that visceral fat was independently associated with kidney growth, that it was a better predictor of faster kidney growth in lean patients than body-mass index, and that having more visceral fat made treatment of ADPKD with tolvaptan less effective.

Introduction

Autosomal dominant polycystic kidney disease (ADPKD) is characterized by the progressive growth of numerous fluid-filled renal cysts, resulting in loss of kidney function in the majority of individuals.¹ There are notable overlapping features and pathways among obesity, dysregulated metabolism, and ADPKD progression.² We have previously reported that overweight and particularly obesity, determined by body-mass index (BMI), are strongly and independently associated with rate of progression in patients with ADPKD who participated in the Tolvaptan Efficacy and Safety in Management of Autosomal Dominant Polycystic Kidney Disease and Its Outcomes (TEMPO) 3:4 trial,³ as well as in the Halt Progression of Polycystic Kidney Disease (HALT) Study A.⁴ However, while BMI correlates with adiposity, it is an indirect surrogate measurement that is considered imprecise.⁵

Adipocytes do not simply act as a fat reservoir, but are active endocrine organs.⁶ Particularly in visceral adipose tissue, adipocytes promote the release of pro-inflammatory cytokines as well as adipokines.⁶ Numerous signaling pathways promoted by adipocytes are also implicated in cystogenesis,^7–9 potentially by promoting a pro-cystogenic milieu. Thus, the association between higher BMI and faster ADPKD progression may be mediated in part by signaling processes from visceral adipose tissue.

The goal of the current study was to quantify visceral adiposity obtained by magnetic resonance imaging (MRI) and its association with ADPKD progression among participants in the TEMPO 3:4 trial. We utilized deep learning to extract a coronal measurement of visceral adiposity and statistical models to evaluate its association with kidney growth (change in TKV), as well as decline in kidney function (estimated glomerular filtration rate [eGFR] slope). We hypothesized that greater abdominal visceral adiposity would independently associate with kidney growth and attenuate the efficacy of tolvaptan in patients with early-stage ADPKD at high risk of rapid progression.

Methods

Surrogate Measurement Definition

In a previous one-year weight loss clinical trial in patients with ADPKD conducted at the University of Colorado Anschutz Medical Campus, 25 paired MRIs were available in both the axial and coronal planes.¹⁰ Additional details are provided in Item S1. Using this feasibility dataset with manual segmentation, a surrogate measurement of axial adiposity using the coronal plane was developed to perform a retrospective analysis on the association of body composition with ADPKD progression in the TEMPO 3:4 clinical trial dataset.^11,12 Manual segmentations of preprocessed volumes were performed using the ITK Snap software, considering visceral fat, subcutaneous fat, and muscle, and using kidney masks (Figure S1). The field of view was calibrated for the best possible resolution of the kidneys, and subcutaneous fat and muscle were rarely acquired. Thus, the study was subsequently focused on visceral fat.

Coronal Measurement Evaluation

To assess the effectiveness of the proposed coronal measurement of the axial body composition, Pearson’s correlation of the percentages of extension of visceral fat in the two available views (three coronal slices and single axial slice) were performed and showed high correlation. Therefore, the segmentation of a three coronal slices slab around the kidney was applied in the next phase of the study. We validated this correlation by performing coronal and axial segmentation of visceral adiposity on baseline images from 55 participants with similar inclusion criteria participating in an ongoing clinical trial (NCT04907799) and obtained comparable results.

Neural Network for Visceral Fat Segmentation

Detailed methodology is provided in Item S1 and Figures S1-S12. The design of the TEMPO 3:4 trial has been described in detail previously.^11,12 We visually classified the TEMPO 3:4 series into three categories (low, medium, and highest visceral adiposity) based on the approximate extent of visceral fat by visual inspection. 328 series were randomly selected from the three categories in a proportionate way for manual segmentation, to define the ground truth dataset to train a neural network for visceral fat identification. The 3-slice slab was extracted from the selected series and annotated. 10% of the series were used as the independent test set, while the remaining ones were divided between the training (70%) and the validation (30%) set. A balanced split was applied, based on the visually classified visceral fat extent (details in Item S1). The training, validation, and test sets thus included 206, 89, and 33 series, respectively. As results were similar using three slices as compared to a single central slice, our subsequent models utilized a single central slice.

Neural Network Architecture

A Unet++ neural network¹³ was used to perform the visceral fat segmentation. It identifies pixels corresponding to visceral fat in the MRI input slices, which are then used to calculate visceral adiposity volume. A pre-trained Unet++ was employed, under a transfer leaning approach, where a model trained on one task is adapted to a new related task. The trained UNet++ reached a 0.86 median Dice Score on the test dataset.

Exploratory Analysis using Unsupervised Methods

Of the 1445 patients in the TEMPO 3:4 trial, 1053 participants with complete data and a BMI ≥18.5 kg/m² were included. Visceral fat volume at the baseline visit was extracted from the trained model using a single central slice. We computed a corrected BMI to remove the contribution of the kidneys total body weight,³ and BMI throughout the text refers to this adjusted BMI. Participants were categorized based on BMI as normal weight (18.5–24.9), overweight (25.0–29.9) or obese (≥30) kg/m².¹⁴ Similarly, we identified three categories based on the tertiles of visceral adiposity volume distribution at the baseline visit, namely low (<41 ml), medium (41 – 83 ml), and high (>83 ml).

We first performed an exploratory data analysis with PatientLens, a software tool developed at OrobixLife supporting the discovery of subtypes in biomedical data to elicit and improve patient stratification.¹⁵

Statistical Modeling

The primary outcome was height-adjusted TKV and the primary predictor variable was visceral adiposity volume, measured at baseline. Standardized MRI scans of the kidneys were performed at baseline and months 12, 24, and 36 (±2 weeks), as well as within 2 weeks of early withdrawal for those not completing the study, as described in detail previously.^11,12

The association of visceral adiposity volume with the longitudinal trajectory of TKV (measured at baseline, 12, 24, and 36 months) was assessed using linear mixed models. Selection of covariates is described in Item S1; time was also included in both models. We also tested for a statistical interaction between visceral adiposity and sex as a predictor of annual percent change in TKV.

To determine whether tolvaptan treatment had a differential effect on the longitudinal trajectory of TKV, we tested the three-way interaction among treatment, time, and visceral adiposity volume in a linear mixed model analysis. We also performed a stratified analysis to compare the differences in annual percent change in TKV across baseline visceral adiposity tertiles according to treatment assignments. .

Similar to our previous BMI analyses,^3,4 we performed multinomial logistic regression models, where the predictor was tertiles of visceral adiposity and the outcome was three categories of the annual percentage of TKV growth (<5%, 5–7%, and ≥7%), calculated as the difference between the logarithmic values of height-adjusted TKV at 36 months and baseline averaged over three years. These ranges were selected to represent low, medium, and high rate of kidney growth based on HALT study A¹⁶ and for comparability to our previous BMI analyses.^3,4 Odds ratios were calculated with the <5% annual TKV growth, lowest visceral adiposity volume tertile serving at the reference. Genotype and plasma copeptin level were added to the final model for the sub-set (n=602) with those variables available.

AIC was computed to evaluate the comparative predictive ability of visceral adiposity in relation to adjusted BMI (as evaluated in our previous publication³) as a predictor of longitudinal trend of TKV in linear mixed models. We also assessed the ability of multiple binary classifiers for either the adjusted BMI or visceral volume as predictors, using the area under the curve (AUC) of the receiver operating characteristic (ROC) curves. AUCs for each classifier were compared with the DeLong’s Test.

We also evaluated the association of visceral adiposity with eGFR slope, as detailed in Item S1.

All statistical analyses were performed using R packages stats v4.3.1, lme4 v1.1–30, lmerTest v3.1–3, and pROC v1.18.0.

Results

Participant Characteristics at Baseline

Among the 1053 participants included in this analysis, age was 39±7 years, 83% (N=875) were White or Caucasian, mean eGFR was 83±22 ml/min per 1.73 m², and annual percent change in TKV was 3.8±4.7%/year. BMI at baseline was 26.2±4.7 kg/m². The median (interquartile range) baseline TKV was 866 (620, 1160) mL. Individuals with greater visceral adiposity volume were more likely to be men and white, have a higher BMI, a higher non-fasting serum glucose, a lower eGFR, a higher blood pressure, a higher plasma copeptin, and a larger TKV at baseline (Table 1).

Table 1.

Baseline characteristics of study participants from TEMPO 3:4 according to tertiles of visceral adiposity volume.

Variable	Tertile 1 (5.0– 40.98 mL) (n=351)	Tertile 2 (41.0 – 83.1 mL) (n=351)	Tertile 3 (83.2 – 399 mL) (n=351)
Age, y	38±7	39±7	40±6
Sex, n (%) Male	108 (31%)	187 (53%)	266 (77%)
Race, n (%) White	285 (81%)	289 (82%)	301 (86%)
Randomization Group (% tolvaptan	210 (60%)	230 (66%)	231 (66%)
BMI, kg/m2	22.9±2.6	25.9±3.9	29.4±4.5
CKD-EPI eGFR, mL/min/1.73m2	88±21	84±22	78±21
Systolic BP, mmHg	128±13	130±15	131±20
Urinary microalbumin, mg/dL	3.6 (2.0, 7.1)	2.5 (1.6, 5.5)	2.5 (1.2, 5.7)
Serum glucose, mg/dL	91±11	94±19	97±16
Plasma copeptin, pmol/L	7.0±11.0	10.9±28.5	11.3±16.2
Mutation class, n (%)
PKD1 truncating	126 (36%)	120 (34%)	116 (56%)
PKD1 non-truncating	61 (17%)	44 (13%)	52 (15%)
PKD2	12 (3%)	24 (7%)	35 (10%)
No mutation detected	4 (1%)	4 (1%)	4 (1%)_
Missing	148 (42%)	159 (45%)	144 (41%)
TKV, mL	1354 (985, 1,821)	1410 (1014, 1961)	1,644 (1243, 2326)
Visceral fat volume, mL	27.6 (19.1, 33.7)	58.6 (49.5, 70.4)	114 (98.1, 146.7)

Data are mean±S.D., %, or median (IQR). CKD-EPI, chronic kidney disease epidemiology collaboration equation; eGFR, estimated glomerular filtration rate; Systolic BP, systolic blood pressure; BMI, body mass-index; TKV, total kidney volume. BMI is calculated from body weight adjusted to remove the contribution of kidney to total weight. Glucose is non-fasting.

^*Plasma copeptin data are missing in 451 participants.

Relation Between Visceral Adiposity and Change in Total Kidney Volume

The three categories of adjusted BMI are clearly visible in the UMAP embedding space, as is the distribution by tertile of visceral adiposity volume (Figures 1A and 1B, respectively); additional UMAP embedding results are presented in Item S2 and Figure S13. The annual percent change in TKV was greater with increasing visceral adiposity, as shown graphically in a spline plot (Figure 2) and across tertiles of baseline visceral adiposity (Figure 3A; tertile 1: 2.3 [2.0, 2.5] %/year, tertile 2: 3.5 [3.2, 3.8]%/year, tertile 3: 5.6 [5.3,5.9] %/year; p<0.01). As expected, baseline visceral volume increased significantly with increasing BMI, with the highest visceral adiposity in individuals who were obese (Figure 4A, Table 1). In the linear mixed model incorporating all available time points where TKV was measured, there was a significant visceral adiposity * time interaction (10.78 [7.17, 14.37], p<2e-¹⁶) (Figure S14). Compared to the lowest visceral adiposity tertile, percent change in TKV from baseline was greater in the middle tertile by 1.52 [0.89, 2.17] % at month 12, 2.2 [1.33, 3.06] % at month 24, and 3.02 [1.91, 4.19] % at month 36. Percent change in TKV from baseline was greater in the highest tertile by 4.7 [3.99, 5.34] % at month 12, 7.39 [6.46, 8.32] % at month 24, and 9.61 [8.41, 10.81] % at month 36.

Figure 1. Embedding spaces for BMI and visceral adiposity volume.

The embedding spaces obtained through PatientLens are shown colored by BMI category (red, obese; yellow, overweight; green, normal weight) and by visceral adipose volume (red, tertile 3 [high]; yellow, tertile 2 [medium]; green, tertile 1 [low]).

Figure 2. Spline plot.

The association of visceral fat volume with annual percent change in total kidney volume is displayed graphically as a spline plot. Knots are set according to tertiles of visceral adiposity. The placebo arm is shown in red and the tolvaptan arm is show in blue.

Figure 3. Annual change in TKV according to baseline visceral adiposity.

Box plots of annual percent change in total kidney volume are shown according to tertiles of baseline visceral adiposity for the entire cohort (A) and divided by tolvaptan and placebo groups (B). Each box describes the first and third quartiles of the data, with the median marked in between; the whiskers extend to the most extreme values of the distribution, up to a maximum of 1.5 times the interquartile range, with values more extreme marked as an outlier. P-values were obtained with a Mann-Whitney-Wilcoxon test with two-side Bonferroni correction. *: 1.00e-02 < p ≤ 5.00e-02, **: 1.00e-03 < p ≤ 1.00e-02, ***: 1.00e-04 < p ≤ 1.00e-03, ****: p ≤ 1.00e-04

Figure 4. Baseline visceral volume according to baseline adjusted BMI categories.

Box plots of baseline visceral adiposity volumes according to baseline adjusted BMI categories (normal weight, overweight, or obese) for the entire cohort (A) and divided by tolvaptan and placebo groups (B). Each box describes the first and third quartiles of the data, with the median marked in between; the whiskers extend to the most extreme values of the distribution, up to a maximum of 1.5 times the interquartile range, with values more extreme marked as an outlier. P-values were obtained with a Mann-Whitney-Wilcoxon test with two-side Bonferroni correction. **** p ≤ 1.00e-04

Similarly, in the fully adjusted multinomial logistic regression model, compared to the lowest tertile of visceral adiposity, the middle tertile had a 1.92 [1.24, 2.95] greater odds and the highest tertile had a 4.78 [3.03, 7.47] greater odds of progressing at a rate of ≥7%/year compared to <5%/year TKV growth, with a similar association when visceral adiposity was considered as a continuous variable (Table 2). Results were similar when excluding extreme visceral adiposity levels >200 mL (Table S1) and when including copeptin and genotype (OR for ≥7%/year vs. <5%/year: 1.94 [1.09, 3.36] (tertile 2 vs. 1) and 3.85 [2.12, 7.02] (tertile 3 vs. 1)).

Table 2.

Associations of visceral adiposity tertiles with annual percent change in total kidney volume.

Endpoint (Annual %Δ in TKV)	Model	5–7% (n=183) vs. <5% (n=606)	≥7% (n=264) vs. <5% (n=606)
Tertile 1 (Visceral adiposity 5.0 – 40.98 mL)	1 (Ref)	1 (Ref)	1 (Ref)
Tertile 2 (Visceral adiposity 41.0 – 83.1 mL)	Unadjusted	1.03 [0.68, 1.56]	1.88 [1.24, 2.83]
	Model 1	1.13 [0.73, 1.74]	2.02 [1.28, 3.16]
	Model 2	1.06 [0.69, 1.63]	1.92 [1.24, 2.95]
Tertile 3 (Visceral adiposity 83.2 – 399 mL)	Unadjusted	1.76 [1.16, 2.64]	4.60 [3.12, 6.83]
	Model 1	1.98 [1.23, 3.16]	5.11 [3.19, 8.17]
	Model 2	1.82 [1.16, 2.86]	4.78 [3.03, 7.47]
Visceral Adiposity as a Continuous Predictor Variable (per 10 unit higher visceral adiposity)	Unadjusted	1.07 [1.04, 1.11]	1.12 [1.09, 1.16]
	Model 1	1.09 [1.04, 1.13]	1.12 [1.09, 1.16]
	Model 2	1.08 [1.04, 1.12]	1.12 [1.08, 1.16]

Reference group is an odds ratio of 1.0. Values are presented as an odds ratio (95% confidence interval). Model 1 was adjusted for age, sex, randomization group, systolic blood pressure, serum glucose, baseline estimated glomerular filtration rate (Chronic Kidney Disease Epidemiology Collaboration equation), and urinary microalbumin. Model 2 was adjusted for age, urinary microalbumin, systolic blood pressure, sex, and hypertension before 35 years.

There was a (non-significant) pattern of greater visceral adiposity at baseline with increasing Mayo Imaging Classification category (Figure S15). When stratifying by baseline Mayo Imaging Classification category, a difference in annual percent change in TKV according to visceral adipose tertiles was evident in categories C-E, but not B (Figure S16). Moreover, there was a statistically significant interaction between visceral adiposity and sex as a predictor of annual percent change in TKV (p<0.01). Results stratified by sex are shown in Table S2. Overall, the association of visceral adiposity with more rapid kidney growth was stronger in females (OR of ≥7% vs. <5% annual change in TKV 6.42 [5.95, 6.94]) for the highest tertile, as compared to males (OR of ≥7% vs. <5% annual change in TKV 4.35 [4.19, 4.51]) for the highest tertile, both in fully adjusted models. Change in visceral adiposity correlated with annual percent change in TKV in both the placebo (R=0.27, p<0.001) and tolvaptan (R=0.32, p<0.001) arm.

Relation Between Visceral Adiposity and eGFR Slope.

In the fully adjusted model 2, the odds of rapid eGFR decline (slope < median −2.77 [−5.32, −1.06]) ml/min/.173m² per year]) were not greater in visceral adiposity tertile 2 (1.36 [0.98, 1.89]) or tertile 3 (1.27 [0.89, 1.84]) (Table S3). Notably, in the fully adjusted models, the association of visceral adiposity as a continuous variable (per 10 unit increase in visceral adiposity) with odds of more rapid decline in eGFR (< median) was statistically significant in females (model 4: 1.06 [1.01, 1.11]) but not males (0.98 [0.95, 1.02]). The association of Mayo Imaging Classification category with eGFR decline was strongest in the highest visceral adiposity tertile (Figure S17).

The Interaction of Adiposity with Tolvaptan Efficacy

The three-way interaction of treatment * time * visceral adiposity was statistically significant (p=0.002) for an outcome of annual change in TKV. The effect of tolvaptan on change in TKV over the study duration, stratified by visceral adiposity tertiles, is shown in Figure 3B. Tolvaptan efficacy was reduced with increasing amount of visceral adiposity (significant three-way interaction). Baseline visceral adiposity volume, stratified by group randomization, is shown in Figure 4B, indicating no differences according to group randomization. Figure 5 graphically displays the three-way interaction; the separation between tolvaptan and placebo regression lines progressively decreases with increasing visceral adiposity volume; this interaction remained statistically significant (p=0.02) after excluding extreme visceral adiposity values ≥200 mL (Figure 5 inset).

Figure 5. Visceral adiposity and annual change in total kidney volume stratified by treatment arm.

Tolvaptan efficacy was reduced with increased visceral adiposity. This figure graphically displays the three-way interaction of treatment * time * visceral adiposity for the outcome of annual change in total kidney volume (TKV); the separation between tolvaptan and placebo regression lines progressively decreases with increasing visceral adiposity volume. The Figure inset removes participants with extreme visceral adiposity values ≥200 mL.

Figure S18 shows eGFR slope in the placebo and tolvaptan arms within each tertile of visceral adiposity; a difference in eGFR slope with tolvaptan was statistically significant in the low and medium, but not high visceral adiposity tertile.

Comparison of Visceral Adiposity to BMI as a Predictor Variable

Adjusted BMI and visceral adiposity volume were significantly correlated (R=0.6, p<2.2e-16). The association of adjusted BMI as a predictor variable with odds of annual change in TKV categories is shown in Table S4, consistent with our prior publication,³ but for the specific cohort in the current analysis. To compare visceral adiposity to adjusted BMI as a predictor of rapid (≥7%) annual change in TKV, the AIC from the mixed model was compared for each predictor. By AIC comparison, the visceral adiposity model was better than BMI for predicting annual change in TKV (a lower AIC of 52378.73 vs. 52923.83). In a comparison of the ROC curves, visceral adiposity did not provide a significant improvement in the classification performance in the whole cohort compared to the adjusted BMI predictor variable (Table 3; Figure S19). However, visceral adiposity significantly improved the classification performance in individuals with a normal BMI, particularly in the tolvaptan randomization arm, potentially due to higher power (2:1 randomization). There were no differences in the ROC curves according to sex.

Table 3.

Area under the curve of receiver operating characteristic curves with adjusted body mass index and visceral adiposity volume as predictors of binary classification of annual change in total kidney volume (≥ 7% compared to <7%).

Drug Cohort	Sub-Group	Predictor	AUC	DeLong’s test Z-Score (p-value)
All participants	None	Corrected BMI	0.66	−0.53 (0.60)
		Visceral Adiposity Volume	0.67
	BMI normal weight	Corrected BMI	0.58	−2.03 (0.04)
		Visceral Adiposity Volume	0.65
	BMI overweight	Corrected BMI	0.57	−0.25 (0.80)
		Visceral Adiposity Volume	0.58
	BMI obese	Corrected BMI	0.54
		Visceral Adiposity Volume	0.54	0.08 (0.93)
Tolvaptan	None	Corrected BMI	0.66	−0.89 (0.37) −0.88 (0.37)
		Visceral Adiposity Volume	0.68
	BMI normal weight	Corrected BMI	0.59	−2.23 (0.03)
		Visceral Adiposity Volume	0.70
	BMI overweight	Corrected BMI	0.53	−0.96 (0.34)
		Visceral Adiposity Volume	0.59
Tolvaptan	BMI obese	Corrected BMI Visceral Adiposity Volume	0.55 0.54	0.14 (0.89)
Placebo	None	Visceral Adiposity Volume	0.68	0.42 (0.67)
		Visceral Adiposity Volume	0.67
	BMI normal weight	Corrected BMI	0.61	−0.56 (0.57)
		Visceral Adiposity Volume	0.63
	BMI overweight	Corrected BMI	0.66	1.00 (0.32) −0.88 (0.37)
		Visceral Adiposity Volume	0.60
	BMI obese	Corrected BMI	0.56	0.43 (0.67)
		Visceral Adiposity Volume	0.51

BMI, body mass index; AUC, Area under the ROC Curve.

Discussion

We have demonstrated that visceral abdominal adiposity quantified by MRI in the coronal plane using a deep learning segmentation model can be used as an index of axial segmentation in patients with ADPKD. Higher visceral abdominal adiposity was independently associated with faster kidney growth in patients with relatively early-stage ADPKD at high risk of rapid progression. Notably, the association was stronger than using BMI as a predictor variable, and there was a significant improvement in risk classification using visceral adiposity instead of BMI in lean individuals. Finally, there was a significant three-way interaction among treatment, time, and visceral adiposity; the efficacy of tolvaptan to slow kidney growth was attenuated in patients with the greatest amount of visceral adiposity.

These findings extend our previous work demonstrating that BMI is independently associated with ADPKD progression^3,4 to focus specifically on abdominal visceral adipose tissue. BMI has inherent limitations, as it is not always reflective of adiposity and is influenced by factors such as muscle mass.⁵ Visceral adiposity quantified in the axial plane (primarily by CT scan) associates with cardiometabolic risk across various cohort studies.^17–21 CT-acquired axial image quantification of visceral adiposity also associates cross-sectionally with cystatin-C based prevalent chronic kidney disease (eGFR <60 ml/min/1.73m²) in the Framingham Heart Study,²² as well as incident chronic kidney disease (eGFR <60 ml/min/1.73m²) and decline in eGFR (30% decrease in eGFR) in the population-based prospective cohort Health ABC.²³ Visceral adiposity by CT is also inversely associated with eGFR in patients with type 2 diabetes.²⁴ Interestingly, in the current study visceral adiposity improved prediction of kidney growth among individuals with a normal BMI, suggesting that visceral adiposity is a more precise predictor of ADPKD progression than BMI and may have potential novel utility in predicting disease progression in lean individuals. Additionally, our results suggest visceral adiposity may explain the association of BMI with kidney growth. Since visceral adiposity is likely already high in overweight and obese individuals, high visceral adiposity may confer particular risk in someone who is lean by BMI, potentially explaining the improved prediction of rapid progression only in this subgroup. This is consistent with evidence of high cardiovascular risk in lean individuals with high central adiposity.²⁵ The strengths of the odds ratios (high visceral adiposity associates more strongly with rapid kidney growth than obesity classified by BMI) further supports this observation.

Notably, visceral adiposity was not significantly associated with kidney function decline in the overall cohort, consistent with our previous study assessing BMI as a predictor of eGFR decline in TEMPO 3:4.³ Participants in the TEMPO 3:4 trial had relatively early-stage ADPKD; thus, it is not surprising that visceral adiposity would be more strongly associated with kidney growth than kidney function decline in this cohort. Indeed, it has been recently highlighted that height-adjusted TKV should be considered in clinical trials as an initial endpoint, with confirmation of an eGFR effect in later CKD-stage patients.²⁶ However, in continuous analyses, it was notable that visceral adiposity was associated with eGFR decline in females but not males. Additionally, baseline eGFR was lowest in visceral adiposity tertile 3, and the association of Mayo Imaging Classification category with eGFR decline was the strongest in the highest visceral adiposity tertile.

Adipocytes do not simply act as a fat reservoir, but are active endocrine organs.⁶ Indeed, visceral adipose tissue produces adipokines more actively than subcutaneous adipose tissue.²⁷ Notably, levels of pro-inflammatory cytokines are elevated in patients with ADPKD. ²⁸ Additionally, epicardial adipose tissue thickness is independently associated with higher levels of the circulating inflammatory marker C-reactive protein in normotensive patients with APDKD and preserved kidney function.²⁹ Abdominal adipose tissue can promote the release of pro-inflammatory cytokines, leading to chronic systemic inflammation, and produce hormone-like factors, including leptin, resistin, and adiponectin.^6,30 Renal sinus fat is negatively correlated with GFR in individuals with type 2 diabetes, supporting the notion that pro-inflammatory cytokines secreted by adipocytes contribute to renal inflammation and dysfunction.³¹ Adipokines and cytokines secreted from adipocytes may promote a pro-cystogenic milieu similar to cancer.^7–9 Additionally, the association between visceral adiposity and kidney growth could also be mediated by insulin resistance^32,33 and/or ADPKD-associated dysregulated cellular kidney metabolism.^2,32

It merits highlighting the lower efficacy of tolvaptan observed with higher visceral adiposity. While a statistical difference in kidney growth between tolvaptan and placebo arms persisted among those with the highest visceral adiposity, the relative efficacy was attenuated as compared to individuals with the lowest visceral adiposity. Additionally, a significant difference in eGFR slope with tolvaptan vs. placebo was also evident in the low and medium, but not high visceral adiposity tertile. Although we did not observe this same significant three-way interaction in our previous study on BMI,³ there was a similar trend, suggesting that BMI may not have been as sensitive as visceral adiposity. There are several possible mechanisms by which tolvaptan efficacy could be reduced with high visceral adiposity. First, with the same absolute dosing, relative dosing of tolvaptan is decreased with greater adiposity. Second, obesity can affect drug metabolism (in part via fatty liver infiltration) and elimination (in part via hyperfiltration);³⁴ however, Otsuka has reported no impact of obesity on tolvaptan pharmacokinetics, as body mass index (range from 17.3 to 90.8 mg/m²) was not a significant covariate in the population pharmacokinetic analysis (unpublished communication). Last, copeptin (a maker of plasma vasopressin), is increased with abdominal adiposity,³⁵ implicated in cyst growth, and a potential biomarker to predict tolvaptan efficacy;³⁶ while levels were higher with increasing visceral adiposity, model results did not change appreciably when including copeptin as a covariate.

There was a significant interaction between sex and visceral adiposity for the outcome change in TKV. While baseline visceral adiposity was greater in males (consistent with literature in the general population ³⁷), high visceral adiposity was on average more strongly associated with kidney growth in females. In particular, visceral adiposity in the highest tertile was very strongly associated with rapid kidney growth in females, and this association was stronger than in males. Similarly, visceral adiposity was only associated with faster decline in kidney function in females. Available evidence suggests that male sex is associated with more severe ADPKD.³⁸ Thus, it is notable that high visceral adiposity may be a particularly important risk factor for rapid progression in females. Interestingly, high visceral adiposity also confers greater cardiometabolic cardiovascular events risk in women than men.^39,40 Additionally, a recent abstract reported a sex difference in the effect of increased fat mass in the Pkd1^RC/RC mouse model of ADPKD; increased PKD severity in response to diet-induced adiposity was observed only in female mice, despite lower levels of adiposity than males.⁴¹ It is possible that sex hormones influence adipose tissue function, contributing to these observations of sexual dimorphism.³⁷

The major strength of our study is the identification of visceral adiposity as a novel factor independently associated with kidney growth in patients with relatively early-stage ADPKD. Additionally, TEMPO 3:4 included a large sample size with a long follow-up and well characterized clinical data. Notable limitations include limited diversity and inclusion of rapid progressors in the TEMPO 3:4 cohort, limiting generalizability of our findings, as well as a lack of mechanistic insight through assessment of circulating markers such as pro-inflammatory cytokines or adipokines. The methodology may be subject to error in ground truth measurement of adiposity, accuracy of the deep learning algorithm, and limitations in reproducibility; results should be confirmed in an external cohort. Finally, the computational demand of the methodology potentially limits clinical translation. However, public sharing of our code, continued advances in machine learning, and the generalizability of obtaining visceral fat in the coronal plane, which is acquired regularly in ADPKD patients in clinical and research setting, are first steps towards minimizing this concern.

In conclusion, visceral adiposity can be quantified by coronally acquired MRI and independently associates with kidney growth, but not eGFR decline, in patients with relatively early-stage ADPKD. High visceral adiposity reduces efficacy of tolvaptan and improves prediction of rapid kidney growth as compared to BMI in normal weight individuals. Quantification of abdominal visceral adiposity by MRI using an automated tool also has implications outside of ADPKD, as previous studies quantifying visceral adiposity have primarily used CT scans and axial images. Future research is needed to verify our findings in another cohort and to understand the mechanisms by which visceral adiposity contributes to ADPKD progression. Additionally, prospective research is needed to determine if a reduction in visceral adiposity slows kidney growth in patients with ADPKD; this is currently being evaluated in an ongoing clinical trial (NCT04907799).

Supplementary Material

Supplementary Material

Item S1. Detailed (full) methods.

Item S2. Additional results description for UMAP embedding.

Figure S1. Preprocessing and manual segmentation of T1 coronal series. Preprocessing of T1 coronal series was performed to extract volumes on which to perform evaluation of a potential surrogate measurement of body composition (A). The central slice of the series was defined as the slice where the kidneys had the largest extension. The slice preceding and following the central slice were also selected, for three slices in total. A slab was extracted around the kidneys in the selected volume by identifying the most extreme points in the y-axis in the kidney contour according to the segmentation of the three highlighted slices and cropping the volume by a margin of 10 pixels above and below the top and bottom extremes, respectively. Next, manual segmentation of the preprocessed volumes was executed, with representative samples shown (B). Regions of visceral fat (green), subcutaneous fat (light blue), and muscle (darker blue) were identified and manually segmented by the Orobix study team.

Figure S2. Schematic of the UNet++ architecture. The UNet++ is a segmentation architecture based on an encoder path, a decoder path, and nested and dense skip connections. In order to bring the semantic level of the encoder feature maps closer to that of the feature maps awaiting in the decoder, in the UNet++ dense convolution blocks are introduced. This architecture has shown top performance on various image segmentation tasks.

Figure S3. Dice loss curves. This figure shows trends of the complementary value of the Dice Score during the UNet++ network training on the training dataset (orange) and validation dataset (blue).

Figure S4. Box plots of Dice distributions in central slice analysis Box plots of the Dice Score distributions of the trained UNet++ performed on the training, validation, and test datasets in the central slice analysis.

Figure S5. Histograms of Dice distributions in central slice analysis. Histograms of the Dice Score distributions of the trained Unet++ performed on the training, validation, and test datasets in the central slice analysis.

Figure S6. Best results on the test set. Images representing the test set series for which the highest Dice Score values have were obtained. The ground truth mask was obtained by manual segmentation and the prediction mask was obtained using the UNet++ network training.

Figure S7. Worst results on the test set. Images representing the test set series for which the lowest Dice Score values were obtained. The ground truth mask was obtained by manual segmentation and the prediction mask was obtained using the UNet++ network training. The series with the lowest results are characterized by particularly limited visceral fat extension. However, the Dice Score is naturally more penalized in these cases: as it is defined, in fact, even small error (e.g., on the contour) has a large impact in terms of Dice if the ground truth pixels are not numerous.

Figure S8. Scatter plot of visceral fat areas. A scatter plot of visceral fat areas calculated by the ground truth masks (obtained by manual segmentation) and those predicted by the neural network.

Figure S9. Bland-Altman plot of visceral fat areas. A Bland-Altman plot of visceral fat areas calculated by the ground truth masks (using manual segmentation) and those predicted by the neural network.

Figure S10. Box plots of Dice distribution on 3 slices. The box plot distributions of Dice Scores of the trained UNet++ performed on the three slices in the training, validation, and test datasets.

Figure S11. Histograms of Dice distribution on 3 slices. Histogram distributions of Dice Scores of the trained UNet++ performed on the three slices in the training, validation, and test datasets.

Figure S12. Scatter plot and a Bland-Altman plot of visceral fat areas using 3 central slices. A scatter plot of visceral fat areas calculated by the ground truth masks (obtained by manual segmentation) and those predicted by the neural network (A) and a Bland-Altman plot of visceral fat areas calculated by the ground truth masks (using manual segmentation) and those predicted by the neural network (B), using 3 central slices.

Figure S13. PatientLens embedding spaces and gradient projection curves. The embedding space obtained through PatientLens are shown colored by sex (A), randomization group (B), age (C), and change in total kidney volume (TKV) at 36 months (D). The gradient projection curve for change in TKV at 36 months as a continuous variable is shown in E and the gradient projection curve for annual change in TKV as a categorical variable is shown in F. The embedding space colored by estimated glomerular filtration rate (eGFR) is shown in G and the gradient project curve for eGFR is shown in H. The embedding space and cluster distribution for urinary microalbumin excretion are shown in I, with graphical quantification of the two identified clusters shown in J.

Figure S14. Visualization of visceral adiposity * time interaction. In the linear mixed model incorporating all available time points where TKV was measured, there was a significant visceral adiposity * time interaction (10.78 [7.17, 14.37], p<2e-¹⁶). The factor time alone increases height-adjusted total kidney volume (htTKV) by 47 units for each year. The visceral adiposity * time interaction indicates that for each increase of 1 standard deviation of visceral adiposity (since the variable was standardized), the effect of time causes an additional increment in htTKV of 10.78 units per year.

Figure S15. Frequency of visceral adiposity tertiles at baseline according to Mayo Imaging Classification. There is a (non-significant) pattern of greater visceral adiposity with increasing Mayo Imaging Classification category at baseline. Please note, due to enrichment of patients at high risk of rapid progression, no participants in TEMPO 3:4 were Mayo Class A.

Figure S16. Annual change in height-adjusted total kidney volume (htTKV) according to Mayo Imaging Classification category within each tertile of visceral adiposity. When stratified by baseline Mayo Imaging Classification category, a difference in annual percent change in TKV according to visceral adipose tertiles was evident in categories C-E, but not B Please note, due to enrichment of patients at high risk of rapid progression, no participants in TEMPO 3:4 were Mayo Class A. Statistical comparisons were performed within each Mayo Imaging Classification category using the Holm method to adjust pairwise t-tests for multiple comparisons. *: p < 5.00e-02, **: 1.00e-02 < p <= 5.00e-02, ***: 1.00e-02 < p <= 1.00e-03, ****: p <= 1.00e-03

Figure S17. Estimated glomerular filtration rate (eGFR) slope according to Mayo Imaging Classification category within each tertile of visceral adiposity. The association of Mayo Imaging Classification category with eGFR decline was strongest in the highest visceral adiposity tertile. Please note, due to enrichment of patients at high risk of rapid progression, no participants in TEMPO 3:4 were Mayo Class A. Statistical comparisons were performed within each tertile by ANOVA. *: p < 5.00e-02.

Figure S18. Estimated glomerular filtration rate (eGFR) slope in the tolvaptan and placebo arms, stratified by baseline visceral adiposity tertile. The difference in eGFR slope with tolvaptan vs. placebo was statistically significant in the low and medium, but not high visceral adiposity tertile. Statistical comparisons were performed within each Mayo Imaging Classification category using pairwise t-tests adjusted using the false discovery rate for multiple comparisons. *: p < 5.00e-02, **: p ≤ 1.00e-02.

Figure S19. Receiver operating curves (ROC) comparing visceral adiposity to adjusted BMI as a predictor of rapid (≥7%) annual change in total kidney volume. ROC curves are presented for the entire cohort (A), stratified baseline body-mass index category (B, normal weight; C, overweight; D, obese). Visceral adiposity significantly improved the classification performance in individuals with a normal BMI, particularly in the tolvaptan randomization arm (E; placebo arm is shown in F; both show the normal weight sub-group).

Table S1. Associations of visceral adiposity tertiles with annual percent change in total kidney volume excluding extreme visceral adiposity values ≥200 mL.

Table S2. Associations of visceral adiposity tertiles with annual percent change in total kidney volume stratified by sex.

Table S3. Associations of visceral adiposity tertiles with odds of faster decline in kidney function (eGFR slope steeper than the median [−2.77 (−5.32, −1.06)] ml/min/1.73m² per year).

Table S4. Associations of BMI with annual percent change in total kidney volume in the same cohort using visceral adiposity as a predictor variable.