AI-enabled chest x-ray potentially detects subclinical diastolic dysfunction in diabesity and monitors its functional responses to GLP-1 or GLP-1/GIP receptor agonists: initial exploration and proof-of-concept

Abstract

Introduction

The evaluation of diastolic function in patients with diabesity (i.e., combined obesity and type-2 diabetes mellitus) undergoing incretin-based therapy could facilitate its early detection and help prevent its progression to heart failure.

Methodology

In this small descriptive study, performed for exploratory and proof-of-concept purposes, a group of 15 diabesity cases with chest x-ray examinations (< 1 month) before and after ≥ 12 months of incretin-based therapy (absent ventricular/valvular dysfunction) were identified. Standard diabesity characteristics (Body Weight (kg), Body Mass Index (kg/m²), and Hemoglobin A1c) and AI-model diagnostic predictions of pulmonary venous hypertension (aka “pulmonary congestion”) [None; Stage 1: vascular distention/redistribution but minimal interstitial edema; or Stage ≥ 2: vascular congestion with ≥ mild interstitial or alveolar edema] by validated AI-enabled chest x-ray staging, representing mean left atrial pressure in diastolic dysfunction, were evaluated pre- and post-therapy.

Results

According to weight-loss response to incretin therapy, cases clustered into equal-sized therapeutic categories as follows: (1) Significant Responders (9–30% decreases); (2) Insignificant Responders (0–2% decreases); and (3) Non-Responders (2–12% increases). Regarding hemoglobin A1c changes: (1) Significant Responders collectively decreased; (2) Insignificant Responders varied; and (3) Non-Responders were largely stable. Pre-therapy, all 15 cases demonstrated AI-enabled chest x-ray pulmonary venous hypertension staging evidence of diastolic dysfunction; post-therapy, 4 improved (especially the cases of greatest weight loss or hemoglobin A1c reduction), 5 were stable, and 6 worsened. Per therapeutic category, AI-enabled chest x-ray signs of functional response were: (1) Significant Responders (all demonstrating unequivocally decreased obesity and improved diabetes) collectively showed stable-decreased dysfunction; (2) Insignificant Responders (all demonstrating at most minimally decreased obesity, but stable-worsening diabetes in most) reflected stable-increased dysfunction in 60%; and (3) Non-Responders (all demonstrating increased obesity, but stable-improved diabetes) showed increased dysfunction in 80%.

Conclusion

AI-enabled chest x-ray pulmonary venous hypertension staging potentially detects background subclinical diastolic dysfunction in diabesity and monitors its functional response to incretin-based therapy. The confirmation (or disproof), and assessments of durability and generalizability, of these preliminary results await a larger more definitive and controlled prospective study on the subject.

Introduction

Obesity [Ob] promotes type-2 Diabetes Mellitus [DM] and the combined “diabesity” has become identified as a major cause of Heart Failure [HF] [1] and a strong focus of pharmacologic management [2, 3]. Especially when found in combination, Ob and DM are myocardial lipotoxic due to increased free fatty acids released from Visceral Adipose Tissue [VAT] being deposited in cardiomyocytes at levels exceeding cellular-storage capacity and insulin resistance-driven demands for fatty-acid oxidation [1]. Consequently, internally leaked deleterious lipid intermediates induce endoplasmic reticulum and mitochondrial injury, inflammatory responses, and cell apoptosis [1]. Associated fibrotic reactions intensify energetics-disrupted myocardial stiffening in causing subclinical Left Ventricular Diastolic Dysfunction [LVDD] [1]. Due to central locations and/or unique compositions, relatively greater lipotoxicity from surrounding Epicardial Adipose Tissue [EAT] and Pericardial Adipose Tissue [PAT] accentuate the lipotoxic effects on LV function, while also exacerbating LVDD by means of surrounding physical restraint of diastolic filling [1].

The benefits of incretin-based therapy with Glucagon-Like Peptide-1 [GLP-1] or combined GLP-1/Glucose-dependent Insulinotropic Polypeptide [GIP] Receptor Agonists [RAs] on the cardiovascular system in the setting of Ob and/or DM are well-recognized [4]. While GLP-1 or GLP-1/GIP RAs improve symptoms related to diabesity-associated HF with preserved Ejection Fraction [HFpEF] [4], their impacts on the asymptomatic pre-clinical phase, including its anticipated progression to HFpEF [5], have not been described. A practical approach to regular evaluation of LV diastolic function in diabesity patients undergoing incretin-based therapy could facilitate early subclinical-LVDD detection and optimize efforts to arrest/reverse progression by medical modification of related HF risk factors [6].

Consequently, in this small descriptive study, performed for exploratory and proof-of-concept purposes, we evaluated LVDD markers provided by Artificial Intelligence [AI]-enabled Chest X-Ray [CXR] staging of Pulmonary Venous Hypertension [PVH] [7] (aka “pulmonary congestion”) in diabesity patients before and after incretin-based therapy. We hypothesized that improved diabesity profiles would be manifested by CXR indicators of improved LV diastolic filling.

Methodology

An Institutional Review Board-approved (including waived consent) search of the shared electronic medical record of the Mayo Clinic enterprise (including its 3 major quaternary medical centers and over 70 smaller facilities located internationally) identified the select group of diabesity patients fortuitously meeting the following strict inclusion criteria:

1. ≥ 12 months of continuous GLP-1 and/or GLP-1/GIP RA therapy.

2. Non-portable digital CXR examinations both within 1 month before diabesity-therapy initiation, as well as within 1 month of completion or after ≥ 12 months of ongoing therapy (if > 1 CXR was applicable, that latest was used).

3. No known or suspected ventricular or valvular dysfunction possibly complicating LVDD assessments.

Following the exclusion of the few patients with conditions potentially confounding further an evaluation of LVDD (1 each: atrial fibrillation; volume-overloading from chronic kidney disease stage 4–5) or a pulmonary vasculature pattern (1 each: secondary pulmonary hypertension; pulmonary fibrosis) [7], the remaining 15 cases constituted the final study group.

In each case, the following diabesity characteristics before and after incretin-based therapy were recorded:

1. Body Weight (kg) and Body Mass Index [BMI] (kg/m²): (https://www.cdc.gov/bmi/adult-calculator/index.html)

2. Hemoglobin A1c [HbA1c]: (Normal = 4.0-5.6%, pre-DM = 5.7–6.4%, and DM ≥ 6.5% https://www.mayoclinic.org/tests-procedures/a1c-test/about/pac-20384643).

In addition, the AI-model prediction probabilities of absent versus present PVH at either Stage 1 (vascular distention/redistribution with minimal interstitial edema) or Stage ≥ 2 (vascular congestion with ≥ mild interstitial or alveolar edema) were determined using AI-enabled CXR PVH staging serving as a surrogate reflecting tendencies for elevation of mean Left Atrial Pressure [mLAP] [8], previously validated for tracking LVDD grades by Doppler Echocardiography [DEcho] [7].

Results

The demographic and background diabesity profiles, as well as subsequent therapies, in the 15 study cases are outlined in Table 1. With decreasing rank ordering of the initial BMI-based classifications (https://www.cdc.gov/bmi/adult-calculator/bmi-categories.html) within 1 week of diabesity therapy initiation, the following unanticipated organization of the cases into three equal-sized BMI subgroups was observed: Cases 1–6 (BMI 41.0-62.6) had Class-3, Cases 7–11 (BMI 35.4–39.8) had Class-2, and Cases 12–14 (BMI 31.4–33.7) had Class-1 Ob, while pre-diabesity Case 15 was considered overweight (BMI 25.3). Within 2 months of the start of therapy, Cases 1–14 demonstrated abnormally elevated HbA1c levels (6.2–12.3%) indicating concurrent DM (12 cases) or pre-DM (2 cases) (https://www.cdc.gov/diabetes/diabetes-testing/prediabetes-a1c-test.html); the pre-diabesity case had a high-normal fasting glucose level (98 mg/dL). Diabesity therapies in the 15 cases spanned 12–46 months between points of CXR monitoring; they included use of a GLP-1 RA (Dulaglutide or Semaglutide) alone (12 cases) or combined GLP-1 RA and GLP-1/GIP RA (3 cases).

Table 1.

Diabesity cases (Sorted by decreasing pre-therapy BMI)

Demographics				Diabesity Profile Before Therapy Initiation				Diabesity Therapy Between CXR Monitoring
Case #	Sex	Age (YO)	Diabesity Components	BMI (kg/m2)	Weight (kg) (< 1 Wk)	HbA1c (Nl 4.0-5.6%) (< 2 Mo)	NT-proBNP** (Neg < 300 pg/mL) (< 2 Wk)	GLP-1 RA	GLP-1/GIP RA (Tirzepatide)	Both
1	F	46	Ob, DM*	62.6	181	7.8	X	Semaglutide 46 Mo (0.25-1.0 mg)	19 Mo (5.0–15.0 mg)	46 Mo Total (Overlapped 19 Mo)
2	F	54	Ob, Pre-DM	50.0	128	6.3	X	Semaglutide 29 Mo (0.25-1.0 mg)	X	X
3	F	81	Ob, DM*	44.9	105	8.1	X	Dulaglutide 101 Mo (0.75–1.5 mg)	X	X
4	F	36	Ob, DM*	42.2	108	7.1	X	Semaglutide 22 Mo (0.25-2.0 mg)	X	X
5	F	50	Ob, DM*	41.2	99	12.3	X	Semaglutide 33 Mo (0.25-1.0 mg)	X	X
6	F	68	Ob, DM*	41.0	109	8.0	X	Semaglutide 54 Mo (0.25-2.0 mg)	X	X
7	F	36	Ob, DM	39.8	111	9.3	X	Semaglutide 13 Mo (0.25-2.0 mg)	10 Mo (5.0-12.5 mg)	13 Mo Total (Overlapped 10 Mo)
8	M	81	Ob, DM*	38.5	122	7.1	X	Semaglutide 24 Mo (0.25-1.0 mg)	X	X
9	M	76	Ob, DM*	35.9	105	9.0	X	Semaglutide 32 Mo (0.25-2.0 mg)	X	X
10	F	64	Ob, DM*	35.4	93	8.6	43	Semaglutide 40 Mo (0.25-2.0 mg)	X	X
11	M	69	Ob, Pre-DM	35.4	111	6.2	X	Semaglutide 24 Mo (0.25-2.0 mg)	12 Mo (5.0-7.5 mg)	36 Mo Total (Sequential)
12	M	82	Ob, DM*	33.7	93	7.1	X	Semaglutide 25 Mo (0.25-2.0 mg)	X	X
13	F	68	Ob, DM*	32.4	81	8.3	X	Semaglutide 12 Mo (0.25-1.0 mg)	X	X
14	M	58	Ob, DM*	31.4	95	11.3	X	Dulaglutide 27 Mo (1.5–4.5 mg)	X	X
15	F	75	Pre-Ob/DM*	25.3	68	FGluc 98 mg/dL	85	Semaglutide 15 Mo (0.25-3.0 mg)	X	X

Units: dL = deciLiters; kg = Kilograms; mg = milligrams; mL = milliLiters; Mo = Months; m² = Meters²; pg = picograms; Wk = Weeks; YO = Years Old

General: BMI = Body Mass Index; DM = type 2 Diabetes Mellitus; F = Female; FGluc = Fasting Glucose; GLP-1 = Glucagon-Like Peptide-1; GIP = Glucose-dependent Insulinotropic Polypeptide; HbA1c = Hemoglobin A1c; M = Male; Neg = Negative; Nl = Normal; NT-proBNP = N-terminal pro-B-type natriuretic peptide; Ob = Obesity; * = added history of Hypertension; ** = concurrent DEcho (< 1 Mo of therapy initiation) had not been performed

The post-therapy diabesity profiles, as well as pre-therapy versus post-therapy AI-enabled CXR PVH staging results, are outlined in Table 2. According to relative (i.e., %) body weight-loss, the cases independently clustered into three distinct equal-sized therapeutic categories as follows: (1) Significant Responders (with 9–30% decreases, consistent with reported target levels of 10–20% weight loss for incretin-based therapy) [9]; (2) Insignificant Responders (with only 0–2% decreases); and (3) Non-Responders (with 2–12% increases).

Table 2.

Diabesity cases after incretin-based therapy (Sorted by decreasing post-therapy weight loss)

Units: kg = kilograms

General: AI = Artificial Intelligence; BMI = Body Mass Index; CXR = Chest X-Ray; GLP-1 = Glucagon-Like Peptide-1; GIP = Glucose-dependent Insulinotropic Polypeptide; HbA1c = Hemoglobin A1c; N/A = Not Applicable; PVH = Pulmonary Venous Hypertension

All Significant Responders also demonstrated HbA1c reductions, including mild (0.5–0.9% point) in 2 cases or moderate (1.0-1.9% point) in 3 cases [10], twice achieving normal levels. Conversely, Insignificant Responders showed varied HbA1c changes, including decreases (1 mild and 1 pronounced ≥ 3%-point) in 2 cases versus moderate increases in 2 cases. Last, HbA1c levels were largely stable in Non-Responders, except for 2 reductions (1 mild and 1 large 2.0-2.9%-point).

All 15 cases exhibited evidence of background pre-therapy LVDD-related mLAP elevations (current and/or chronic) by AI-enabled CXR determinations of PVH [7], either at Stage 1 (11 cases) or Stage ≥ 2 (4 cases) [Table 2]. With diabesity therapy, 4 cases demonstrated CXR evidence of decreased mLAP elevation based on: (1) Decreased likelihood (i.e., highest AI-model probability reduced by ≥ 0.10) at the same PVH stage alone (2 cases); or (2) Prominent decreased PVH likelihood causing downward transitioning from Stage ≥ 2 to Stage 1 (2 cases: Significant Responder Case 3 with the greatest weight loss of 30% (Fig. 1a), and Insignificant Responder Case 14 with the greatest HbA1c decrease). In contrast, 6 cases indicated mLAP elevation based on: (1) Increased AI-model likelihood of the same PVH stage alone (2 cases); or (2) Prominent increased PVH likelihood causing upward transitioning from Stage 1 to Stage ≥ 2 (2 Insignificant Responders and 2 Non-Responders Fig. 1b). The remaining 5 cases, including 3 Significant Responders, demonstrated evidence of stable mLAP elevation with therapy.

Fig. 1.

In both (a) and (b), AI-model attention maps (right) generated from AI-enabled PVH staging of original CXRs (left) demonstrate changing attention distributions and intensities (yellow-red colors) both pre- and post- incretin-based therapy with GLP-1 or GLP-1/GIP RAs. The Significant Responder (Case 3) in (a) shows decreased overall attention activity in the lungs, especially upper-lung, reflecting prominent decreased PVH likelihood causing downward transitioning from Stage > 2 to Stage 1. On the other hand, the Non-Responder (Case 10) in (b) shows increased overall and upper-lung activity, representing prominent worsening PVH likelihood causing upward transitioning from Stage 1 to Stage > 2

Thus, per therapeutic category, the following pattern of LVDD-related functional responses to diabesity therapy was observed (Table 2): (1) Significant Responders (all unequivocally demonstrating weight loss approaching or surpassing target levels [9] while retaining Ob-level BMI values of 31.4–55.1, in combination with improved DM) collectively showed evidence of stable-to-decreased mLAP elevations (especially Case 3); (2) Insignificant Responders (all securing Ob-level BMI values of 31.1–49.0, but stable-worsening DM in most) reflected stable-to-increased mLAP elevations in the majority (60%); and (3) Non-Responders (all demonstrating advancing Ob levels, but stable-improved DM) showed evidence of progressing mLAP elevations in the large majority (80%, the exception being the initially pre-diabesity case).

Discussion

This small descriptive study from a single large international multi-site healthcare enterprise was performed for exploratory and proof-of-concept purposes; it provides preliminary evidence of background subclinical LVDD associated with a range of diabesity in patients apparently representing the asymptomatic pre-clinical phase of the HF spectrum [1, 11]. Provision of this insight was shown to be feasible using a validated AI-enabled CXR method for recognizing PVH stages which track LVDD grades determined using DEcho while presumably reflecting tendencies for mLAP elevations, likely intermittent due to fluctuating loading conditions [7, 8, 12].

In this study, AI-enabled CXR PVH staging capabilities also produced preliminary evidence supporting stabilized-to-decreased mLAP elevations following distinct improvements in Ob and/or DM markers (elevated body weight and/or HbA1c level, respectively) after incretin-based therapy with GLP-1 or GLP-1/GIP RAs [4]. On the other hand, relatively stable-to-increased body weight and/or worsening DM-status despite therapy appeared to enable LVDD progression, possibly eventually towards HFpEF [5].

The reported experience with this specific application of AI-enabled CXR PVH staging resulted in pilot data motivating further development of the initial hypothesis and justifying a future larger more definitive and controlled prospective study to confirm (or disprove) these initial results.

Limitations and considerations

Nevertheless, this descriptive study has several limitations discussed as follows:

1. The retrospective approach to data collection is a recognized but unavoidable limitation.

2. The small size of the final 15-case study group was an uncontrollable consequence of the lack, until now, of any known evidence endorsing systematic CXR-based evaluations of diabesity patients during GLP-1 and/or GLP-1/GIP RA therapy; however, it conceivably could become a practical (e.g., affordable, widely available, easily integrated) diabesity-care enhancement.

3. Data analysis remained semi-quantitative because of risk of the small size of our study group providing spurious and imprecise results impacting an apparent “statistical significance” of a relationship [13, 14]. Nevertheless, this pilot investigation revealed an interesting pattern of therapeutic responses that, at this time, encourages further investigation of this application of AI-enabled CXR PVH staging for the detecting subclinical LVDD in diabesity and monitoring its functional responses to incretin-based therapy.

4. While there is considerable heterogeneity (related to case profiles and incretin-based therapy administrations) within this small group, the fact that a pattern of functional responses was observed suggests that this heterogeneity provided valuable relative real-world implementation advantages [15]. Further consideration of this heterogeneity was considered to be beyond the scope of this study.

5. Whether or not the initial BMI levels affected the diagnostic value of AI-enabled CXR PVH staging is not clear, however it is notable that representatives of all three initial BMI subgroups were subsequently represented in all three therapeutic categories. This included the five cases with highest initial BMI values having become a Significant Responder (N = 3), an Insignificant Responder (N = 1), or a Non-Responder (N = 1).

6. Concurrent DEcho data was not available for further validation of AI-enabled CXR PVH staging results related to this specific application. In the absence of symptoms in these 15 cases apparently representing the pre-clinical phase (either Stage A/at-risk or Stage B/pre-HF) [11], DEcho had not been consistently performed, and, when done, examinations were not in synchrony with CXR examinations at the incretin-therapy endpoints. However, in a previous study, AI-enabled CXR PVH staging was validated against DEcho LVDD grading, including in Grade 0 and Grade 1 LVDD (unlike in Healthy normals) which revealed increasing evidence of PVH attributed to rising tendencies for intermittent mLAP elevations from characteristically varying loading conditions with or without chronic pulmonary vascular alterations [7, 12]. Nevertheless, it remains possible that AI-enabled CXR PVH staging could misrepresent the current status of LV diastolic filling at any particular point in time. Hence, re-evaluation of the relationship between PVH staging by AI-enabled CXR and LVDD grading by DEcho (based on new recommendations [16]) will be incorporated into the planned prospective study, while still recognizing that DEcho may be inconclusive regarding LVDD (e.g., 14% for 2016 recommendations [17]) even when using the newer versions [16, 18].

7. Last, this study did not incorporate imaging of VAT, EAT, or PAT deposit changes from therapy.

Conclusion

AI-enabled CXR PVH staging, presumably reflecting LVDD-related mLAP elevation tendencies, potentially both detects background subclinical LVDD characterizing diabesity and monitors functional responses (improvement, stabilization, or progression) to incretin-based therapy. However, the confirmation (or disproof) of these favorable preliminary results, and assessments of their durability and generalizability, await a planned larger more definitive and controlled prospective study on the subject.

Abbreviations

AI

Artificial Intelligence

BMI

Body Mass Index

CXR

Chest X-Ray

DEcho

Doppler Echocardiography

DM

Type-2 Diabetes Mellitus

EAT

Epicardial Adipose Tissue

GIP

Glucose-dependent Insulinotropic Polypeptide

GLP-1

Glucagon-Like Peptide-1

HbA1c

Hemoglobin A1c

HF

Heart Failure

HFpEF

HF with preserved Ejection Fraction

LVDD

Left Ventricular Diastolic Dysfunction

mLAP

Mean Left Atrial Pressure

Ob

Obesity

PAT

Pericardial Adipose Tissue

PVH

Pulmonary Venous Hypertension

RAs

Receptor Agonists

VAT

Visceral Adipose Tissue

Author contributions

Individual author contributions are as follows: RDW: Supervised all aspects from Concept development to Short Report submission. MD: Concept confirmation, CXR & AI-model processing, Short Report review/revision. BSE: Concept confirmation, case data mining, AI-model processing, Short Report review/revision. GL: Concept confirmation, case data analysis, Short Report review/revision. All authors read and approved the final version of the Short Report for submission.

Funding

This work was entirely supported internally by the Department of Radiology, Mayo Clinic Florida; no external funding was used.

Data availability

No established databases were utilized in support of this study. Complete image datasets are not made publicly available for non-federally funded research due to Mayo Clinic institutional restrictions. However, the high-resolution CXR images used in this study may be available to interested researchers upon request from the Center for Augmented Intelligence in Imaging of the Mayo Clinic Florida (contact: erdal.barbaros@mayo.edu).

Declarations

Ethics approval and consent to participate

The needed compiling and analysis of the focused data mined from the Mayo Clinic enterprise electronic medical record, with waived requirement for patient consenting, was performed with prior approval from the Mayo Clinic Institutional Review Board. No Generative-AI technology was used in the performance of this research or the preparation of this Short Report.

Consent for publication

The use of the de-identified or anonymized data in this Short Report was approved by the Mayo Clinic Institutional Review Board.

Authors’ information

RDW: Cardiovascular imager since mid-1980’s with MS in imaging-AI / Medical Director-Center for Augmented Intelligence in Imaging at Mayo Clinic Florida. MD: Computer engineer with AI subspecialization. BSE: Electrical and computer engineer with data-mining and AI subspecialization / Technical Director-Center for Augmented Intelligence in Imaging at Mayo Clinic Florida. GL: Chair-Cardiovascular Medicine at Mayo Clinic Florida with heart failure subspecialization.

Supplementary information

Specific technical details and validation procedures related to the AI-enable CXR methodology applied in this study can be found at https://pubmed.ncbi.nlm.nih.gov/41174036/.

Competing interests

Other than an application filed by the Mayo Clinic (including White, Demirer, and Erdal) for a provisional patent to protect the used AI methodology while we gain real-world experience such as this, the authors have no potential competing interests to declare.