Artificial Intelligence-Driven Longitudinal Quantification of Technetium Pyrophosphate Uptake in Cardiac Amyloidosis: Correlation with Multimodality Imaging and Outcomes

Abstract

Background:

Transthyretin cardiac amyloidosis (ATTR-CM) is an increasingly recognized cause of heart failure (HF) in older adults. Several therapies for ATTR-CM are now available, with more currently in development. As such, there is an increasing need for methods to assess response to therapy. We evaluated the associations between serial 99m-Technetium pyrophosphate (^99mTc-PYP) deep learning measurements with changes in other imaging parameters and clinical outcomes.

Methods:

We included patients with a diagnosis of ATTR-CM and at least two ^99mTc-PYP studies followed through the Amyloidosis Program of Calgary. Patients underwent laboratory testing, echocardiography, and cardiovascular magnetic resonance (CMR) unless contraindications were present. ^99mTc-PYP images were quantified using our previously developed deep learning methodology including assessment of cardiac pyrophosphate activity (CPA) and volume of involvement (VOI).

Results:

In total 85 patients were included, with median population age 79 (interquartile range 72 – 84) and 76 (89%) male patients. In patients on therapy, there was a reduction in VOI (median 100 to 51, p<0.001), CPA (median 165 to 81, p<0.001), native T1 (median 1399 to 1380, p=0.029), and extracellular volume (median 52 to 50, p=0.031) during a median time of 369 days (interquartile range 365 – 516) between scans. There was a modest correlation between change in CPA with change in native T1 (ρ=0.376, p=0.009). After adjusting for age, treatment, and CPA at follow-up, an increase in CPA during follow-up was also associated with increased risk (adjusted HR 2.31 per SD increase, 95% CI 1.28 – 4.17, p=0.005).

Conclusions:

Serial ^99mTc-PYP quantitation has modest correlations with other measures of disease burden including native T1. Changes in these measures were associated with risk of cardiovascular death or HF hospitalization, suggesting that the serial measurements may be clinically meaningful surrogate endpoints.

Graphical Abstract

INTRODUCTION

Transthyretin cardiac amyloidosis (ATTR-CM) is an increasingly recognized cause of heart failure (HF) in older adults, with technetium-99m pyrophosphate (^99mTc-PYP) established as a highly accurate tool for establishing the diagnosis (1). Hybrid SPECT/CT imaging facilitates robust methods to quantify myocardial radiotracer activity. Scully et al. demonstrated high diagnostic accuracy for quantitation using standardized uptake values from regions placed manually on SPECT/CT images(2). These metrics have high diagnostic accuracy(3), and are correlated with measures of fibrosis from cardiovascular magnetic resonance (CMR) imaging(4). Therefore, these quantitative metrics may provide valuable diagnostic and prognostic information to support clinical decision making.

Deep learning has emerged as an effective method to provide these quantitative assessments without tedious manual segmentations. We demonstrated the feasibility of quantifying these measurements in a fully automated fashion using deep learning with excellent diagnostic accuracy(5). In our analyses, cardiac pyrophosphate activity and volume of involvement at baseline were also associated with risk of HF hospitalization or cardiovascular death(5). This approach could provide an ongoing assessment of disease severity as a measure of response to therapy. This is an increasingly important clinical need given the emergence of multiple therapeutic options for ATTR-CM(6–9). However, it has not been conclusively demonstrated whether these measurements could be used to assess response to therapy(10).

We quantified serial ^99mTc-PYP imaging parameters using deep learning in a cohort of patients with ATTR-CM. We then evaluated changes in these measures in response to therapy as well as correlations with other clinical, imaging, and outcome-based measures of disease progression.

MATERIALS & METHODS

Study Population

In this retrospective study we identified patients who underwent at least two ^99mTc-PYP SPECT/CT imaging at the University of Calgary (Calgary, Alberta, Canada) between March 2018 and January 2025. This study was reviewed and approved by the Institutional Review Board at the University of Calgary (REB19-1448).

Clinical Data

Demographics and past medical history were determined at the time of the first ^99mTc-PYP. Diagnosis of ATTR-CM was based on the presence of 1) endomyocardial biopsy positive for ATTR, or 2) diffuse myocardial uptake on SPECT/CT imaging after exclusion of monoclonal protein, and 3) evidence of infiltrative cardiomyopathy on echocardiography or cardiac magnetic resonance imaging. All patients with a confirmed diagnosis underwent pathologic gene variant testing to confirm ATTR subtype s(11, 12). Patients were followed by our institution’s multidisciplinary amyloidosis program, the Amyloidosis Program of Calgary and underwent annual clinical evaluation including laboratory (troponin and N-terminal pro-B-type natriuretic peptide [NT-proBNP]), ^99mTc-PYP imaging and echocardiographic evaluation. Patients without contraindications also underwent annual cardiovascular magnetic resonance (CMR) imaging. Patients who received an ATTR-CM stabilizing agent or ATTR silencing drug for at least half of the follow-up time were considered as “treated”. Treatment decisions were based on several clinical factors as well as changes in insurance coverage as previously described(13). National amyloid center stage was determined at baseline(14).

CMR Imaging Protocols and Interpretation

CMR examinations were performed using 3.0T systems (Prisma or Skyra, Siemens Healthineers, Erlangen, Germany). The majority of patients with serial imaging were imaged using the same system for both baseline and follow-up imaging (43/45, 96%). Standard Cine SSFP imaging was performed in sequential short axis views and 2-, 3- and 4-chamber long-axis views for the evaluation of LV volume, function and mass. Native T1 mapping was performed in three short-axis views (basal, mid and apical) using a MOLLI pulse sequence (15). LGE imaging was performed using a phase-sensitive inversion recovery (PSIR) gradient echo pulse sequence as previously described (16). Specifically, the imaging was achieved in sequential short- and long-axis views 10-minutes following the intravenous administration of 0.15 mmol/kg Gadovist (Bayer Inc., Mississauga, Canada).

Quantitative image analysis was performed using commercially available software (cvi⁴², Circle Cardiovascular Inc., Calgary, Alberta, Canada) at the time of clinical reporting. Short axis cine images underwent semi-automated contour tracing to obtain the LV end-diastolic volume (EDV), end-systolic volume (ESV), ejection fraction (EF) and LV mass. LV mass was determined at end-diastole with the inclusion of the dominant papillary muscles. All volume and mass measures were indexed to body surface area (BSA). LGE was visually coded using a standardized reporting interface (cardioDI^™, Cohesic Inc., Calgary). For each segment of the standard AHA 17-segment model(17), the presence and transmural extent (0: no LGE, 1: 1–25% thickness, 2: 26–50%, 3: 51–75%, 4: 76–100%) of LGE was scored. Global LGE burden was estimated as: total segmental score / 68 (total possible score), as previously described (18).

PYP Imaging protocols

All subjects underwent imaging on a GE NM Discovery 870 scanner. Patients received 740 Mbq (20 mCi) of ^99mTc-PYP, with planar images obtained at 3-hours post injection over 5–8 minutes duration, with a heart-centered field of view. SPECT images were acquired at 3 hours with 10% energy window and 128 × 128 matrix. CT attenuation correction (CTAC) imaging was performed during end-expiratory breath hold with no gating, in helical mode with a slice thickness of 5-mm, tube voltage of 120 kVp and 20 mA, using a 512×512 matrix for both scanner systems. Planar myocardial radiotracer retention was quantified using heart to contralateral (HCL) ratio at the time of clinical interpretation from the 3-hour images(11). Planar images at 3-hours were also graded visually using Perugini grading(1).

Deep Learning Quantification

We used our previously developed approach for quantification, with measurements performed on attenuation corrected images(5). In our previous work, CPA and VOI demonstrated consistently excellent diagnostic accuracy and therefore the present analysis focused on these measurements(5). The background region of interest was a shrunk left atrium segmentation. VOI is quantified as the total volume of activity above threshold (defined as the maximum activity of the background)(3). CPA is quantified as the total activity (integrating volume and intensity) above threshold, divided by mean background activity(3). While many other quantifications are possible, we chose to focus on two measurements to avoid significant type 1 error given the sample size. We compared manual and automated quantitation of CPA and VOI on baseline and follow-up imaging. For manual quantification, the background region of interest was placed in the left atrium and the target region of interest encompassed the LV myocardium and LV cavity since it is not feasible to manually segment LV myocardium alone.

Outcomes:

We evaluated associations with the combined clinical outcome of cardiovascular death or HF hospitalization. Hospitalizations for HF were ascertained from electronic medical records (19). Cardiovascular mortality was ascertained from electronic medical records and included death related to HF, myocardial infarction, and sudden death(20). Follow-up was censored at the last time that follow-up status could be verified.

Statistical Analysis

Categorical variables were summarized as numbers (proportions) and compared with a Chi-square or Fisher exact test as appropriate. Continuous variables were summarized as mean (standard deviation) and compared with a student’s t-test if normally distributed, and with a Wilcoxon rank sum test if not. Change in laboratory and imaging variables over time was evaluated using Wilcoxon signed-rank tests. Differences in change between groups (treated vs. untreated) were evaluated using Wilcoxon rank-sum tests. Correlations between baseline, follow-up, and change in variables were assessed using Spearman’s correlation coefficients, with Bonferroni-corrected p-values to adjust for multiple testing. Associations between follow-up imaging and laboratory values with the combined outcome of cardiovascular death or HF hospitalization were assessed using Fine Gray competing risk models, with non-cardiovascular death as the competing risk. Models included age, sex, and use of targeted therapy at any point during follow-up. Change in laboratory or imaging variables was also assessed if the association between follow-up value was significant. Imaging and laboratory variables were modeled as continuous values normalized to standard deviation to allow for comparability between values. The follow-up time started at the time of the follow-up ^99mTc-PYP scan to account for immortal time bias. All statistical tests were two-sided, with p<0.05 considered statistically significant. All analyses were performed with STATA version 13 (StataCorp, College Station, Texas).

RESULTS

Patient Population

In total, 85 patients were included in the study with population characteristics outlined in Table 1. The median age was 79 (interquartile range 72 – 84) and 76 (89%) patients were male. The majority of treated patients received tafamadis (n=71, 96%), with the remaining patients receiving patisiran (n=3, 4%)

Table 1:

Population characteristics at baseline. ATTR-transthyretin cardiac amyloidosis CAD - coronary artery disease, IQR – interquartile range, NYHA – New York Heart Association.

	All patients (N=85)	Untreated n=11	Treated n=74	p-value
Age, median (IQR)	79 (72 – 84)	84 (81, 85)	78 (71, 83)	0.016
Male, n(%)	76 (89%)	7 (64%)	69 (93%)	0.003
NYHA Class				0.72
1	7 (9%)	0 (0%)	7 (10%)
2	51 (65%)	7 (70%)	44 (59%)
3 or 4	20 (26%)	3 (30%)	17 (25%)
Hereditary ATTR-CM	5 (6%)	1 (9%)	4 (5%)	1 (9%)
Medical History, n(%)
Hypertension	53 (62%)	6 (55%)	47 (64%)	0.57
Diabetes	13 (15%)	3 (27%)	10 (14%)	0.24
CAD	16 (19%)	5 (45%)	11 (15%)	0.015
Atrial Fibrillation	39 (46%)	3 (27%)	36 (49%)	0.18
Carpal tunnel syndrome	33 (39%)	3 (27%)	30 (41%)	0.40
Spinal Stenosis	19 (22%)	3 (27%)	16 (22%)	0.67
Peripheral neuropathy	35 (41%)	7 (70%)	28 (41%)	0.080

Change in Imaging Parameters Over Time

The median time between baseline and first follow-up ^99mTc-PYP was 369 days (IQR 365 – 516). Change in ^99mTc-PYP quantitative measures over time are shown in Figure 1. Patients on ATTR therapies had a significant decrease in CPA (median 165 to 81, p<0.001) and VOI (median 100 to 51, p<0.001) from baseline to first follow-up. The median change in CPA was −64 (IQR −8 to −162), with 57 patients (77%) having a decrease over time. The median change in VOI was −36 (IQR 0.3 to −102), with 54 patients (73%) having a decrease over time. In patients not on therapy, CPA (228 to 206, p=0.186) and VOI (105 to 127, p=0.155) did not significantly change. The median change in CPA was 17 (IQR −28 to 50), with 3 patients (27%) having a decrease over time. The median change in VOI was 12 (IQR −16 to 26), with 3 patients (27%) having a decrease over time.

Figure 1:

Change in technetium-99m pyrophosphate measures from baseline to follow-up.

Echocardiograms were available in all patients and occurred at a median of 44 days (IQR 26 – 61) before baseline ^99mTc-PYP and a median of 28 days (IQR 14-44) before follow-up ^99mTc-PYP. Baseline CMR was available in 55 patients and occurred a median of 28 days after ^99mTc-PYP (IQR 71 before to 63 after) and follow-up CMR was available in 54 patients at a median of 11 days after follow-up ^99mTc-PYP (IQR 10 before to 53 after). Change from baseline to follow-up measures, stratified by treatment, for Echocardiogram, CMR, and ^99mTc-PYP is shown in Table 2. In addition to changes in VOI and CPA in patients on therapy, there was a reduction in native T1 (median 1400 to 1387, p=0.005), and extracellular volume (median 50 to 49, p=0.035). The median change in native T1 was −18 (IQR −45 to 5), with 27 patients (60%) having a decrease over time. The median change in ECV was 0 (IQR −3 to 0), with 11 patients (31%) having a decrease over time.

Table 2:

Imaging characteristics. Continuous variables are presented as median (interquartile range).

	Untreated (n=11)			Treated (n=74)
	Baseline	Follow-up	p-value	Baseline	Follow-up	p-value
PYP
Planar Grade			0.317			0.049
0	0 (0%)	1 (1%)		0 (0%)	1 (1%)
1	1 (9%)	4 (5%)		3 (4%)	3 (4%)
2	2 (18%)	19 (22%)		10 (14%)	16 (22%)
3	8 (73%)	61 (72%)		61 (82%)	54 (73%)
HCL Ratio	1.8 (1.25, 2)	1.6 (1.25, 1.98)	0.928	1.71 (1.51, 1.89)	1.58 (1.43, 1.8)	0.006
VOI	105 (34, 170)	127 (57, 231)	0.155	100 (60, 176)	51 (26, 107)	<0.001
CPA	228 (76, 367)	206 (110, 339)	0.594	165 (99, 329)	81 (45, 187)	<0.001
Biomarkers
Troponin	63 (23, 145)	56 (30, 145)	0.423	45 (31, 59)	49 (30, 69)	0.064
NT-proBNP	5282 (3062, 14522)	4582 (3050, 10592)	0.313	1730 (767, 2921)	1726 (927, 2853)	0.152
Echocardiogram
LVEF	50 (38, 57)	50 (39, 57)	0.888	50 (42, 59)	48 (40, 57)	0.001
LVEDD	3.9 (3.5, 4.3)	4.2 (3.5, 4.4)	0.582	4.4 (4, 4.9)	4.4 (4.1, 4.8)	0.251
LVESD	2.9 (2.7, 3.3)	3.0 (2.7, 3.3)	0.958	3.3 (2.8, 3.8)	3.2 (2.8, 3.6)	0.742
IVSD	1.8 (1.6, 2)	1.7 (1.5, 1.8)	0.288	1.6 (1.4, 1.7)	1.6 (1.5, 1.8)	0.697
LPWD	1.8 (1.5, 1.9)	1.7 (1.4, 1.8)	0.660	1.5 (1.4, 1.7)	1.5 (1.4, 1.7)	0.105
CMR*
LVEF	52 (44, 57)	54 (42, 56)	1.000	51 (47, 60)	50 (44, 58)	0.071
LVMI	116 (97, 128)	106 (104, 121)	0.052	84 (75, 101)	89 (72, 104)	0.670
RVEF	44 (36, 49)	39 (33, 46)	0.281	53 (46, 61)	52 (45, 55)	0.111
Native T1	1433 (1424, 1476)	1438 (1412, 1476)	0.095	1400 (1351, 1450)	1387 (1336, 1425)	0.005
% LGE	73 (60, 75)	75 (70, 75)	0.317	50 (40, 68)	50 (37, 63)	0.202
ECV	66 (63, 69)	64 (62, 65)	0.317	50 (46, 58)	49 (45, 56)	0.035

^*-Cardiovascular magnetic resonance (CMR) variables limited to patients with both baseline and follow-up imaging (n=45), with paired late gadolinium enhancement (LGE) measurements available in 40 patients and paired extracellular volume (ECV) measurements in 35 patients. CPA – cardiac pyrophosphate active, HCL – heart to contralateral, IVSD – interventricular septal dimension, LVEDD – left ventricular end-diastolic dimension, LVESD – left ventricular end-systolic dimension, LVMI – left ventricular mass index, NT-proBNP – N-terminal pro-B-type natriuretic peptide, VOI – volume of involvement.

The correlation between baseline imaging measures is shown in Figure 2. There was a modest correlation between CPA and native T1 (ρ=0.514, p=0.003). The correlation between follow-up measures is shown in Supplemental Figure 1. Follow-up CPA and VOI were correlated with native T1 (ρ=0.517, p=0.008 and ρ=0.517, p=0.008 respectively). The correlation between change in measurements is shown in Figure 3. There were modest correlations between change in CPA with change in native T1 (ρ=0.376, p=0.009) and change in VOI with change in native T1 (ρ=0.335, p=0.021). Change in HCL ratio was not correlated with change in CMR parameters. There was excellent agreement between deep learning and manual segmentations with intraclass correlation coefficients of 0.979 for CPA and 0.963 for VOI. Bland-Altman plots are shown in Supplemental Figure 2 and 3.

Figure 2:

Correlation between baseline variables. CMR – cardiovascular magnetic resonance, CPA – cardiac pyrophosphate active, ECV – extracellular volume, HCL – heart to contralateral, IVSD – interventricular septal dimension, LVEDD – left ventricular end-diastolic dimension, LVESD – left ventricular end-systolic dimension, LGE – late gadolinium enhancement, LVMI – left ventricular mass index, NT-proBNP – n-terminal pro-B-type natriuretic peptide, VOI – volume of involvement.

Figure 3:

Correlation between change in variables. CMR – cardiovascular magnetic resonance, CPA – cardiac pyrophosphate active, ECV – extracellular volume, HCL – heart to contralateral, IVSD – interventricular septal dimension, LVEDD – left ventricular end-diastolic dimension, LVESD – left ventricular end-systolic dimension, LGE – late gadolinium enhancement, LVMI – left ventricular mass index, NT-proBNP – n-terminal pro-B-type natriuretic peptide, VOI – volume of involvement.

Associations with Outcomes

During a median follow-up of 584 days (IQR 263 – 894) after the follow-up ^99mTc-PYP, first events included 12 patients hospitalized for HF and 16 patients who died (11 cardiovascular deaths, 5 non-cardiovascular deaths). Associations between follow-up variables and cardiovascular death or HF hospitalization are outlined in Table 3. After adjusting for age and use of targeted therapies, VOI (adjusted HR 1.59 per SD increase, 95% CI 1.26 – 2.00, p<0.001), CPA (adjusted HR 1.70 per SD increase, 95% CI 1.39 – 2.09, p<0.001), and extracellular volume (adjusted HR 2.55 per SD increase, 95% CI 1.05 – 6.17, p<0.001) from follow-up imaging were associated with an increased risk of cardiovascular death or HF hospitalization. After adjusting for age, sex, treatment, and CPA at follow-up, an increase in CPA was also associated with increased risk (adjusted HR 2.31 per SD increase, 95% CI 1.28 – 4.17, p=0.005). Similar results were seen for VOI (adjusted HR 2.29 per SD increase, 95% CI 1.44 – 3.64, p<0.001), and ECV (adjusted HR 2.38 per SD increase, 95% CI 1.13 – 5.01, p=0.023). While follow-up values for NTproBNP and troponin were associated with outcomes, neither change in NTproBNP (adjusted HR 0.91 per SD increase, 95% CI 0.66 – 1.25, p=0.720) nor troponin (adjusted HR 0.87 per SD, 95% CI 0.60 – 1.26, p=0.459) were associated with outcomes. Similar associations were seen after also adjusting for national amyloid center stage at baseline, with an increased risk of cardiovascular death or HF hospitalization associated with increase in CPA (adjusted HR 2.06 per SD increase, 95% CI 1.01 – 4.23, p=0.048), increase in VOI (adjusted HR 1.62 per SD, 95% CI 1.01 – 2.62, p=0.046), and increase in ECV (adjusted HR 2.41 per SD, 95% CI 1.23 – 4.70, p=0.010).

Table 3:

Associations between follow-up laboratory and imaging variables with cardiovascular death or heart failure hospitalization. Follow-up time started at the second (follow-up) imaging test. All hazard ratios (HR) reflect the risk associated with one standard deviation increase from Fine Gray competing risk models. Adjusted for age, sex, use of targeted therapies as a time-varying covariate. CPA – cardiac pyrophosphate active, HCL – heart to contralateral, IVSD – interventricular septal dimension, LVEDD – left ventricular end-diastolic dimension, LVESD – left ventricular end-systolic dimension, LVMI – left ventricular mass index, NT-proBNP – n-terminal pro-B-type natriuretic peptide, VOI – volume of involvement.

	Standard deviation	Adjusted HR (95% CI)	p-value
HCL Ratio	0.35	1.18 (0.84 – 1.66)	0.346
VOI	81.7	1.59 (1.26 – 2.00)	<0.001
CPA	153	1.70 (1.39 – 2.09)	<0.001
Troponin	51.6	1.91 (1.48 – 2.48)	<0.001
NTproBNP	3741	2.16 (1.55 – 3.02)	<0.001
Echocardiogram LVEF	11.3	0.91 (0.59 – 1.41)	0.672
LVEDD	0.62	1.16 (0.67 – 2.00)	0.606
LVESD	0.75	1.09 (0.73 – 1.65)	0.664
IVSD	0.27	1.20 (0.75 – 1.91)	0.455
LPWD	0.27	1.23 (0.65 – 2.34)	0.519
CMR LVEF	10.5	1.30 (0.70 – 2.42)	0.400
LV Mass Index	20.3	0.99 (0.61 – 1.62)	0.974
RVEF	24.9	0.78 (0.44 – 1.35)	0.376
Native T1	61.1	1.57 (0.76 – 3.24)	0.227
% Myocardial LGE	18.9	1.42 (0.91 – 2.22)	0.120
Extracellular Volume	8.8	1.97 (1.05 – 3.72)	0.036

Case Examples

Two case examples are shown in Figure 4 and Figure 5. The first case demonstrates images from an 83-year-old male patient treated with tafamadis. During the first year of follow-up, there was a significant reduction in both CPA (315 to 198) and VOI (275 to 107), with corresponding improvements in native T1 (1430 to 1420) and echocardiographic LVEF (26% to 47%). The second case was a 77-year-old male patient who was not initially started on therapy due to concerns regarding cognitive decline. During 18 months of follow-up there was progressive increase in both CPA (375 to 467) and VOI (170 to 299).

Figure 4:

Case example from an 83-year-old male patient treated with tafamadis. After the first year of follow-up, there was a significant reduction in both cardiac pyrophosphate activity (CPA) and volume of involvement (VOI) with corresponding improvements in native T1 and echocardiographic left ventricular ejection fraction (LVEF).

Figure 5:

Case example from a 77-year-old male patient who was not initially started on therapy due to concerns regarding cognitive decline. After 18 months of follow-up there was an increase in both cardiac pyrophosphate activity (CPA) and volume of involvement (VOI).

DISCUSSION

We evaluated the utility of serial ^99mTc-PYP deep learning measurements in assessing response to therapy. We demonstrated that in the absence of treatment, ^99mTc-PYP quantitative measures did not significantly change over time, while they tended to decrease in patients on therapy. CPA and VOI did not have significant correlations with biochemical or echocardiographic measures of disease severity. However, CPA and VOI did demonstrate moderate correlations with native T1 and ECV at baseline and follow-up as well as correlations with change in native T1. Lastly, follow-up ^99mTc-PYP quantitative measures, as well as change in ^99mTc-PYP quantitative measures over time, were associated with risk of HF hospitalization of cardiovascular death. These results suggest that serial ^99mTc-PYP deep learning measurements could potentially play a role in monitoring response to therapy in patients with ATTR-CM.

The management of patients with ATTR-CM has dramatically shifted over the last 5 years. In our center, as with many centers, access to ATTR stabilizers has transitioned to near universal coverage as the standard of care(21). Many patients with hereditary ATTR-CM are now treated with gene silencers, with potential future therapies including gene editing approaches(22) and amyloid depleters(23). With this changing therapeutic landscape and the availability of multiple therapeutic options, there is increasing need for methods to accurately evaluate disease progression and response to therapy.

We evaluated several possible methods for evaluating disease progression and response to therapy. In our analysis, we did not identify significant changes in laboratory biomarkers or echocardiographic measures of disease severity in patients on therapy. In contrast, we identified significant changes in ^99mTc-PYP measurements, as well as native T1 and ECV, in patients on therapy. Similarly, Vijayakumar et al. identified changes in ^99mTc-PYP quantitative measures over time in 23 patients on therapy but did not identify significant changes in laboratory or echocardiographic measures(10). In our analysis, we also identified moderate correlations between CMR and ^99mTc-PYP quantitative measures both as baseline measures as well as change in measurements. It is not clear whether these changes reflect amyloid regression or plaque stabilization, but studies using amyloid specific radiotracer suggest it may reflect the latter(24). Volumetric heart to lung ratios for ^99mTc-PYP uptake decrease in patients receiving eplontersen(25), with reductions in standardized uptake values demonstrated with tafamadis therapy using ^99mTc-PYP (26) or diphosphono-1,2-propanodicarboxylic acid(27). Similar to our results, Patel et al. recently demonstrated that many CMR measures of disease severity progress in patients not on therapy and tend to stabilize in patients on therapy(28). While there was no difference overall in ECV among patients receiving patisiran, those who did experience a reduction in ECV were less likely to experience all-cause mortality compared to patients who experienced an increase in ECV(28).

Given uncertainty with how to best assess disease severity at any specific point in time, we also evaluated associations with outcomes. Laboratory biomarkers, including troponin and NT-proBNP, are included in the Mayo clinic staging of ATTR-CM(29) and were associated with cardiovascular events in our analysis. However, changes over time were not associated with events, likely reflecting the strong prognostic value of baseline troponin and NT-proBNP levels. CMR measurements have also been used to assess response to therapy in prior studies(30). In our analysis, both baseline ECV and change in ECV were associated with incidence of HF hospitalization or cardiovascular death, similar to other recent reports(28), suggesting the CMR may play in important role in disease monitoring. Follow-up CPA and VOI, as well as change in CPA and VOI, were associated with cardiovascular risk in our analysis. Overall, these results suggest that cardiac imaging, with ^99mTc-PYP and CMR may play complementary roles in following response to therapy. In particular, worsening CPA, VOI, or ECV in the setting of therapy for ATTR-CM may be an indication to consider alternative agents or combination therapy.

In contrast to other potential measures of disease progression ^99mTc-PYP demonstrates a few important properties. Firstly, ^99mTc-PYP is almost always performed to establish a diagnosis of ATTR-CM meaning baseline measurements are near universally available. Compared to CMR, ^99mTc-PYP imaging is relatively accessible enabling its use as a serial measure. Furthermore, manual ^99mTc-PYP quantitation has similar interobserver variability as quantitation of the H/CL ratio. We demonstrated excellent correlations between expert manual and deep learning segmentations. However, there was a bias towards lower values with deep learning segmentations which may relate to differences in technique, with the manual segmentations also including activity above threshold in the LV cavity. Deep learning-based segmentation removes variability related to manual placement of background and target regions of interest, potentially further improving reproducibility. This has particularly important implications for serial assessment as regions will be placed in precisely the same manner with each study. Importantly, there have been several approaches for deep learning segmentation, which may not all require CT attenuation correction imaging. Bhattaru et al. utilize deep learning to segment the cardiac silhouette from CT attenuation correction imaging, similar to our approach(31). However, studies have used direct identification of abnormal CARI uptake on planar imaging(32, 33). While the same approach could be challenging to apply to SPECT imaging for diagnosis (given variability in blood pool signal), it may be feasible in the setting of disease monitoring (where uptake is more uniform). Lastly, it is worth noting that amyloid-specific radiotracers have many of the same potential strengths with the added benefit of known mechanisms of binding(34).

Our study has a few important limitations. We quantified all SPECT/CT images at 3-hours using ^99mTc-PYP and the results may not extrapolate to centers only performing imaging at 1 hour(35), or using other radiotracers(36). Given sample size limitations we were not able to evaluate variation in change in ^99mTc-PYP separately in patients managed with different drug classes or fully evaluate the incremental prognostic value of ^99mTc-PYP quantification in addition to other risk scores. Additionally, the analysis regarding change in ^99mTc-PYP in patients not receiving therapy is underpowered. Furthermore, future studies are needed to determine whether the association between change in ^99mTc-PYP is also seen in a cohort of patients who all receive treatment. The mechanism for ^99mTc-PYP uptake is not fully understood and we do not have data regarding histopathologic correlations with the proposed AI measures. The changes we identified may be a reflection of change in amyloid fibril volume or stability. However, the exact mechanism may be less important than its ability to predict clinical outcomes. Our AI-based approach for ^99mTc-PYP was initially developed using data from a single center, which is also the center in the present work. Further validation in external sites, with a variety of imaging protocols, is an important next step which could be facilitated through large registries(37). Lastly, not all measurements were performed at the same time periods given some variability in clinical follow-up over time. While this study design limits our ability to provide annualized estimates of change, it did allow us to assess long-term clinical follow-up.

CONCLUSIONS

Serial ^99mTc-PYP quantitation has modest correlations with other measures of disease burden including ECV and LVEF. However, reduction in CPA was associated with risk of cardiovascular death or HF hospitalization suggesting that the changes may still be clinically meaningful.

NEW KNOWLEDGE GAINED

This study is one of the largest to date to follow changes in laboratory and multimodality imaging quantitative biomarkers in patients with ATTR-CM, including patients receiving and not receiving therapies. We demonstrated that multiple measures change over time, with modest correlation between ^99mTc-PYP and CMR measures. However, both CMR and ^99mTc-PYP measures (including change over time) were predictive of the combined outcome of HF hospitalization or cardiovascular death.

CLINICAL PERSPECTIVE

Our results suggest that AI-based quantitation of ^99mTc-PYP uptake may play a role in the multimodality assessment of patient-level response to ATTR-CM therapies.