AI-based integrated body composition assessment from cardiac SPECT CT attenuation maps and its association with mortality ​

Jirong Yi, PhD^a, Krishna K. Patel, MD^b, Anna M. Marcinkiewicz, MD, PhD^a,c, Aakash Shanbhag, MSc^a,d, Robert J.H. Miller, MD^a,e, Wenhao Zhang, PhD^a, Aditya Killekar, MSc^a, Mark Lemley, BSc^a, Jianhang Zhou, MSc^a, Joanna X. Liang, MPH^a, Giselle Ramirez, BS^a, Valerie Builoff, BSc^a, Terrence D. Ruddy, MD^f, Andrew J. Einstein, MD, PhD^g, Attila Feher, MD, PhD^h, Edward J. Miller, MD, PhD^h, Daniel S. Berman, MD^a,  Damini Dey, PhD^a, Piotr J. Slomka, PhD^a ​

^a Cedars-Sinai Medical Center, USA ^b Icahn School of Medicine at Mount Sinai, USA ^c National Medical Institute of the Ministry of the Interior and Administration, Poland ^d University of Southern California, USA ^e University of Calgary, Canada ^f University of Ottawa Heart Institute, Canada ^g Columbia University Irving Medical Center and New York-Presbyterian Hospital, USA ^h Yale University, USA ​

Body Composition Analysis

• CT attenuation correction (CTAC)
• Correct for soft-tissue attenuation
• Provide additional anatomic information

​

​

​

• Body composition analysis
• Reflect the amount and distribution of body tissues
• Provide biomarkers for phenotyping and risk stratification

​

CTAC

Segmentation

Body Composition Analysis

Miller et al., European Heart Journal 2025

Proven independent prognostic values of multiple body composition metrics in cohort with high event rate (33%)

Body Composition Analysis

Yi et al., Lancet Digital Health 2025

Proven independent prognostic values of multiple body composition metrics in cohort with low event rate (6%)

To develop AI model for comprehensive body composition assessment

​

• Integration of 19 body composition metrics

​

• Integration of 6 different tissues

​

• Prognostic significance evaluation of integration

Aims

Cohort creation

• Computed tomography for attenuation correction (CTAC)
• Epicardial adipose tissue (EAT)

See Yi et al., Lancet Digital Health 2025 for more details

Cohort creation

4 sites from REFINE SPECT registry

Subcutaneous adipose tissue

Skeletal muscle (SM)

Intramuscular adipose tissue

Bone

Visceral adipose tissue

Epicardial adipose tissue

Evaluation for Mortality Prediction

​

• Discriminative prediction: area under curve

​

• Prognostic prediction: hazard ratio

Body Composition Segmentation and Quantification [1,2]

Methods

Integrated Body Composition Model Development and Validation

​

• 5-fold cross validation for training
• 10-fold testing for evaluation

[1] Miller et al., European Heart Journal 2025

[2] Yi et al., Lancet Digital Health 2025

19 quantification from T5-T11

• Mean attenuation (6 tissues)
• Attenuation standard deviation (6 tissues)
• Volume index (all tissues)
• Volume ratio of muscle to adipose tissue

Variables for Adjustment

19

Metrics

Score between 0 and 1

Results: quantification metrics

Visceral adipose tissue (VAT

Bone

Skeletal muscle (SM)

Subcutaneous adipose tissue (SAT)

Intramuscular adipose tissue (IMAT)

Epicardial adipose tissue (EAT)

See Yi et al., Lancet Digital Health 2025 for more details

Results: feature importance

VAT – visceral adipose tissue, SAT – subcutaneous adipose tissue, EAT – epicardial adipose tissue, IMAT – intramuscular adipose tissue, SM – skeletal muscle, SD – standard deviation

• VAT attenuation is the most important feature

​

• For VAT, SAT, IMAT: attenuation is most important feature

​

• For bone, SM, and EAT: volume index is the most important feature

Results: body composition phenotyping

Visceral adipose tissue (VAT

Bone

Skeletal muscle (SM)

Subcutaneous adipose tissue (SAT)

Intramuscular adipose tissue (IMAT)

Epicardial adipose tissue (EAT)

Maximally selected rank test-based cutoff 0.57 was used for defining high-risk (body composition score ≥ 0.57) and low-risk (body composition score < 0.57)

Results: improvement over individual metric

*

*

*

*

Prediction power of automated and integrated body composition score (BCS) for all-cause mortality prediction

​

* p<0.0001

SM – skeletal muscle, VAT – visceral adipose tissue, SM2AT – ratio of skeletal muscle volume index to adipose tissue volume index

Body composition score (BCS)

Unadjusted HR: 4.8 [4.09, 5.64], p<0.0001

Adjusted HR: 3.09 [2.6, 3.69], p<0.0001

Results: prognostic value of integrated assessment

Imaging quantification data (3):

stress total perfusion deficit, left ventricular ejection fraction, coronary artery calcium score

​

Clinical data (8): age, body mass index, sex, family coronary artery disease history, diabetes, dyslipidemia, hypertension, smoking

11 covariates for adjustment

Maximally selected rank test-based cutoff: 0.57 [Abbas et al., Nature Communications 2023]

Cox Regression Hazard Ratio Results

EAT - Epicardial Adipose Tissue, IMAT - Intramuscular Adipose Tissue, SAT - Subcutaneous Adipose Tissue, SM - Skeletal Muscle, VAT - Visceral Adipose Tissue

Results: high-risk case

Age: 68 years

Sex: male

Body mass index: 17.67 kg/m²

Died at follow up of 0.4 years

-65 Hounsfield unit

-86 Hounsfield unit

-57 Hounsfield unit

-45 Hounsfield unit

181 Hounsfield unit

31 Hounsfield unit

42 Hounsfield unit

32 Hounsfield unit

54 Hounsfield unit

1.6

46 Hounsfield unit

214 Hounsfield unit

148 cm³/m²

498 cm³/m²

119 cm³/m²

351 cm³/m²

19 cm³/m²

28 cm³/m²

32 Hounsfield unit

Results: low-risk case

Age: 52 years

Sex: male

Body mass index: 27.1 kg/m²

No death during entire follow up of 6.2 years

43 Hounsfield unit

-85 Hounsfield unit

1.2

31 Hounsfield unit

-62 Hounsfield unit

30 Hounsfield unit

-70 Hounsfield unit

291 Hounsfield unit

189 Hounsfield unit

-91 Hounsfield unit

40 Hounsfield unit

47 cm³/m²

37 Hounsfield unit

56 Hounsfield unit

26.4 cm³/m²

1040 cm³/m²

537 cm³/m²

268 cm³/m²

337 cm³/m²

• Automatic and effective integration of metrics of multiple body composition tissues from CT is feasible

Conclusions

Acknowledgement: Supported in part by Grants R01HL089765 and R35HL161195 from the National Heart, Lung, and Blood Institute/National Institutes of Health (NHLBI/NIH) (PI: Piotr Slomka).

• Integrated body composition assessment can improve all-cause mortality prediction over individual metric

• Integrated body composition assessment can provide added prognostic value over clinical and perfusion variables

Appendix

1

2

10

…

Test data fold 1

(N=992)

Developing data fold 1

(N=8926)

…

Equally-partitioned into ten folds (N=9918)

1

2

5

3

4

Equally-partitioned

into five folds (N=8926)

…

Validation

data fold (1,1)

(N=1785)

Training data fold (1,1)

(N=7141)

XGBoost

model training

XGBoost

model evaluation

Config 1

Config 2

…

Candidate hyperparameter configurations

AUC on validation data fold 1

Repeat for all validation folds

Repeat for all hyperparameter configurations

Select the hyperparameter configuration with highest average AUC

Repeat for all 10 testing data folds

(A)

Integration Algorithm Design – Hyperparameter Tuning

AUC averaged on five validation data folds

Config 1

Test data fold 1

(N=992)

…

Equally-partitioned into ten folds (N=9918)

20%

80%

20%

Validation data 1

(N=1785)

Training data 1

(N=7141)

XGBoost

model training

XGBoost

model evaluation and selection

Selected hyperparameter configurations for fold 1

Config

Selected XGBoost model for test data fold 1

Developing data fold 1

(N=8926)

Randomly split into training and validation data (N=8926)

Repeat for all testing data folds

A selected XGBoost models

(B)

Integration Algorithm Design – Model Training and Selection

1

2

10

…

80%

Integration Algorithm Design – Model Inference and Testing

1

2

10

Selected integration model from development data fold 1

…

Selected integration model from development data fold 10

…

Selected integration model from development data fold 2

Concatenate predictions of all testing data folds

…

Prediction on testing data fold 1

Prediction on testing data fold 10

…

Prediction on testing data fold 2

Prognostic evaluation

Testing data folds