Study design and participants
This retrospective case-control study followed institutional guidelines and was approved by the Memorial Sloan Kettering Cancer Center (MSK) Institutional Review Board, protocol number 17-581. The requirement for informed consent was waived because the study involved secondary analysis of previously acquired PET-CT imaging and clinical data, posed minimal risk to participants, and could not be conducted practically with individual informed consent. All patient data was anonymized before use.
We previously established a cohort of patients with BAT by analyzing 18F-FDG-PET-CT scans performed at MSK from June 2009 to March 2018.4. Patients who were positive for BAT on at least one analysis were referred to as BAT+while the remaining patients were grouped into the BAT category−. To expand this cohort, we created a new Extension BAT cohort following a similar strategy. We reviewed all 18F-FDG-PET-CT scans performed at MSK on patients aged 18 years and older from April 2018 to July 2024. Individuals who were positive for BAT on at least one scan were referred to as BAT.+while the remaining patients were grouped into the BAT category− (Extended data, Fig. 1a and Supplementary Fig. 2). For each patient in both groups, an index scan was designated: in patients with BAT reported on at least one 18F-FDG-PET-CT scan, the first scan reporting BAT was defined as the index scan. In patients without BAT on any scan, the first 18F-FDG-PET–CT scan served as the index scan. We collected demographic data, BMI, outdoor temperature, and relevant ICD-9 and ICD-10 codes corresponding to cardiometabolic diseases.4 (1 year after index analysis) (Supplementary Table 8) of the institutional EHR. Average outdoor temperatures in New York were obtained from the United States National Weather Service. A first logistic regression analysis was carried out on BAT patients+ against B.A.T.− groups belonging exclusively to the BAT Extension cohort to establish associations between the presence of BAT and cardiometabolic diseases. As the association of the presence of BAT with reduced prevalence of cardiometabolic diseases was confirmed in the BAT expansion cohort (extended data, Fig. 1b), we proceeded to merge the established BAT cohort with the BAT expansion cohort to create an extended BAT cohort (extended data, Fig. 1a). If a patient was present in both cohorts, data from the established BAT cohort were used.
Data collection and PSM
To create a balanced study cohort derived from the expanded BAT cohort, BAT patients+ and BAT− groups were matched using PSM. Propensity scores were estimated using a non-parsimonious multivariate logistic regression model with BAT status (presence or absence) as the dependent variable and age, sex, BMI, and temperature at the time of analysis as covariates. The matching was performed using a greedy nearest neighbor matching algorithm without replacement, with a thickness width set to 0.2 and a matching ratio of 1:2. This resulted in a final study cohort of 26,302 (8,897 BAT+ and 17,405 BAT−). Balance diagnostics were performed to assess the success of the matching procedure using standardized mean differences for all variables included in the propensity score model before and after matching. Post-matching standardized mean differences between −0.10 and 0.10 indicated adequate balance between BAT+ and BAT− groups for all measured covariates, supporting the validity of the PSM model for causal inference.
We collected relevant clinical data for each patient as previously described.4. Briefly, for cancer site, we included the cancer-specific diagnosis code (ICD-0) listed closest to the index analysis date (Supplementary Table 1). Data referring to cancer treatment were collected if administered within 90 days prior to the index analysis and classified as before.4 (Supplementary Table 2). Relevant ICD-9 and ICD-10 codes corresponding to cardiometabolic diseases4 (Supplementary Table 8) and all other codes were extracted up to 1 year after the index analysis in patients in our PSM cohort. Codes were selected to represent non-pediatric, non-obstetric disorders in adults, as explained below.
For laboratory values, clinical indices, and imaging studies, we extracted data available up to 1 year after the index analysis in our PSM cohort. Analyzes were carried out in BAT+ and BAT− patients with available values. To identify individual pharmacological agents associated with the presence of BAT, we extracted medications from the home medication list and/or medications with an active pharmaceutical prescription in the EHR of each patient belonging to our PSM cohort. We included only medications listed up to 1 month before the index analysis for each patient. We excluded pharmacological agents with unlikely systemic effects, such as those administered topically, ophthalmically, nasally, and otically. Antimicrobials were excluded because they are generally prescribed in short courses. Medications classified as “as needed” and herbal supplements were excluded. Low frequency medications (present in less than 40 patients) were excluded. A final list of drugs included in the analyzes can be found in Supplementary Table 9. These drugs were also grouped by general categories as described in Supplementary Table 9.
Selection of diagnostic codes for unbiased associations with BAT
To explore unbiased associations between BAT and other diseases, we extracted EHR diagnosis codes (ICD-9 and ICD-10) listed up to 1 year after the index analysis of patients belonging to our PSM cohort. To select which codes to extract, we grouped ICD-9 and ICD-10 codes based on their first three letters and/or numbers. ICD-10 codes belonging to the general groups: A, B, C, D, E, F, G, H, I, J, K, L, M, N, and R were included because they group medical conditions most likely to be found in an adult population with non-pediatric, non-obstetric cancer. A similar rationale was used for the selection of ICD-9 codes. A total of 1,157 ICD-10 codes and 626 ICD-9 codes were queried for extraction from the EHR.
After code extraction, to avoid redundancy, we merged the ICD-9 codes with their corresponding ICD-10 codes based on manual review of each code. ICD-10 code nomenclature was retained for downstream analysis in cases where an ICD-10 and ICD-9 code were merged. ICD-9 codes that did not have a corresponding ICD-10 code were retained with their nomenclature. This left us with a total of 1,209 codes between the individual ICD-9 codes and the merged ICD-9 and ICD-10 codes.
To further filter diagnosis codes for analyses, we excluded low-frequency codes (present in fewer than 40 patients) and those referring to a cancer diagnosis (Supplementary Table 1), given that all patients already had an assigned cancer site. This left us with a final list of 511 diagnostic codes. Given the large number of codes, we initially performed a screening analysis using univariate logistic regression analysis to identify unadjusted associations of 511 with the presence of BAT. A total of 104 codes were significantly associated with BAT status in this analysis. Subsequently, we analyzed each of these 104 codes using multivariate regression analysis to assess their individual association with BAT, while controlling for age, sex, BMI, ambient temperature, self-reported race and ethnicity, smoking history, family history of cardiovascular disease, cancer site, cancer stage, cancer treatment, and general outpatient medication groups (Extended Data, Fig. 1a).
Machine learning approach to quantify BAT activity
The 3DF workflow19 used PET and CT images in DICOM format. Using information from the DICOM header regarding image orientation and spatial positioning of the patient, the PET image was resampled to match the dimensions of the CT image by linear interpolation. The PET and CT images were then converted to .nrrd file format and processed by nnU-net. The nnU-net model produced a mask prediction where each pixel value corresponds to an anatomical site: 1 for cervical, 2 for supraclavicular, 3 for mediastinal, 4 for paraspinal, 5 for axillary, and 6 for abdominal regions. The mask predictions were interpreted by 3DF and the contours were generated. These contours were downsampled to match the dimensions of the original PET image, and the original PET measurements for each anatomical site were exported to an Excel master file. The 3DF process was run in batch mode, with multiple instances running simultaneously to efficiently segment a large cohort. We used this method to quantify the metabolic activity of BAT in 8,709 BAT patients.+ group from our cohort. A small portion of patients (n= 188) in the BAT+ The group could not be quantified due to differences in field of view between CT and PET images.
Machine learning approach to quantify SAT and VAT area
The 3DF19.60 The abdominal adipose tissue workflow was performed on axial CT images at the level of the L3–L4 intervertebral disc. The target slice was identified automatically using TotalSegmentator61 vertebral segmentation, the analysis section being defined as the midpoint between the lower extent of the L3 mask and the upper extent of the L4 mask. The external contour of the body on the selected slice was derived by thresholding voxels above -500 Hounsfield units, followed by gap filling, removing small components, retaining the largest connected component, and morphological closure. External abdominal circumference was then calculated from the resulting body boundary using DICOM pixel spacing. Adipose tissue was quantified using a hybrid attenuation and AI-based method. Large voxels were first identified in the body contour using an attenuation range of -195 to -45 Hounsfield units, and then these voxels were divided into SAT and TVA using TotalSegmentator tissue-like bin masks. The external abdominal circumference (in centimeters), the SAT area (in square centimeters), the VAT area (in squares…
Gn Health
