PET connectomes can identify long COVID patients from those without persistent complaints
Maria Ferrandez, Nelleke Tolboom, Denise Visser, Bart Van berckel, Robert Schuit, Bert Windhorst, Ronald Boellaard and Sandeep Golla
Introduction
Total body PET imaging with 18F-DPA-714 provides a unique opportunity to study inflammation-related alterations across the whole body in patients with long COVID. This project aimed to quantify organ-to-organ uptake connectivity in long COVID patients compared to healthy controls, i.e. patients without persistent complaints, using standardized uptake values (SUV) and network-based correlation analyses.
Methods
A total of 34 dynamic EARL2 PET/CT scans from Long COVID patients (22) and post COVID patients without complaints (control group, n=12) were included in this study. Ten anatomical structures were segmented from the CT images: myocardium, lungs, kidneys, liver, spleen, stomach, brain, bone (skull and spine), subcutaneous fat and skeletal muscle.
Any misalignment between CT and PET due to motion or physiological changes were manually corrected. Standard uptake values corrected for injected activity and body weight (SUVBW) were extracted from the last three PET frames (30–60 min). Group-level connectivity was computed as the organ-by-organ Pearson correlation matrix across individuals (10×10). The reference network (refNET) was defined as the pearson correlation matrix derived from the control group.
To assess individual deviations, we constructed perturbed networks (ptbNET) by adding each patient to the control group and recalculating correlations. An individual residual network (resNET) was defined as the difference between the ptbNET and the refNET, in order to isolate the contribution of a specific patient to the control group. Network node strength (STR) was calculated as the sum of absolute residual correlations per organ.
Group differences in SUV and STR were assessed using the non-parametric Mann–Whitney U test. P-values were considered significant at p < 0.05. Both SUV and STR were used in predictions models for the classification of subjects. The models were trained using leave-one-out cross validation (LOOCV). The area under the curve (AUC), accuracy, confidence intervals and confusion matrices were used to assess the features performance.
Results
Significant SUV differences between Long COVID patients and controls were observed, most prominently in the heart (p<0.005, based on mann-whitney statistical testing), with additional differences in the lungs (p<0.01), bone (p<0.01), kidneys (p<0.05) and spleen (p<0.05).
Connectivity analysis revealed that STR was also significantly altered, especially in the heart (p<0.005), lungs (p<0.05), and kidneys(p<0.05), indicating disrupted inter-organ coupling. Individual resNETs were different between groups, with a greater amount and more intense organ connections in patients compared to controls (Figure 1).
Classification models combining SUV and network features demonstrated good discriminatory performance. A logistic regression model using heart SUV and heart STR achieved an AUC of 0.86 [95% CI: 0.72–0.96] with a sensitivity of 60%, a specificity of 93% and a false-positive rate of 0.06.
A random forest model incorporating all 10×10 residual correlations and SUV values from the seven most discriminative organs (heart, spleen, lungs, kidneys, brain, liver, and skull/spine) achieved an AUC of 0.80 [95% CI: 0.64–0.95], a sensitivity of 67%, a specificity of 85% and a false positive rate of 0.15.
Conclusions
This study demonstrates that Long COVID is associated with distinct alterations and systemic connectivity changes detectable with 18F-DPA-714 total body PET. The heart was found to be a key organ showing both abnormal SUV and altered network strength, with additional contributions from kidneys and lungs.
Combining SUV and connectivity features provides a promising approach to reveal abnormal inter-organ interactions in Long COVID and supports the potential clinical utility of network-based PET analysis.
Web | DOI | Journal of Nuclear Medicine | Abstract
Maria Ferrandez, Nelleke Tolboom, Denise Visser, Bart Van berckel, Robert Schuit, Bert Windhorst, Ronald Boellaard and Sandeep Golla
Introduction
Total body PET imaging with 18F-DPA-714 provides a unique opportunity to study inflammation-related alterations across the whole body in patients with long COVID. This project aimed to quantify organ-to-organ uptake connectivity in long COVID patients compared to healthy controls, i.e. patients without persistent complaints, using standardized uptake values (SUV) and network-based correlation analyses.
Methods
A total of 34 dynamic EARL2 PET/CT scans from Long COVID patients (22) and post COVID patients without complaints (control group, n=12) were included in this study. Ten anatomical structures were segmented from the CT images: myocardium, lungs, kidneys, liver, spleen, stomach, brain, bone (skull and spine), subcutaneous fat and skeletal muscle.
Any misalignment between CT and PET due to motion or physiological changes were manually corrected. Standard uptake values corrected for injected activity and body weight (SUVBW) were extracted from the last three PET frames (30–60 min). Group-level connectivity was computed as the organ-by-organ Pearson correlation matrix across individuals (10×10). The reference network (refNET) was defined as the pearson correlation matrix derived from the control group.
To assess individual deviations, we constructed perturbed networks (ptbNET) by adding each patient to the control group and recalculating correlations. An individual residual network (resNET) was defined as the difference between the ptbNET and the refNET, in order to isolate the contribution of a specific patient to the control group. Network node strength (STR) was calculated as the sum of absolute residual correlations per organ.
Group differences in SUV and STR were assessed using the non-parametric Mann–Whitney U test. P-values were considered significant at p < 0.05. Both SUV and STR were used in predictions models for the classification of subjects. The models were trained using leave-one-out cross validation (LOOCV). The area under the curve (AUC), accuracy, confidence intervals and confusion matrices were used to assess the features performance.
Results
Significant SUV differences between Long COVID patients and controls were observed, most prominently in the heart (p<0.005, based on mann-whitney statistical testing), with additional differences in the lungs (p<0.01), bone (p<0.01), kidneys (p<0.05) and spleen (p<0.05).
Connectivity analysis revealed that STR was also significantly altered, especially in the heart (p<0.005), lungs (p<0.05), and kidneys(p<0.05), indicating disrupted inter-organ coupling. Individual resNETs were different between groups, with a greater amount and more intense organ connections in patients compared to controls (Figure 1).
Classification models combining SUV and network features demonstrated good discriminatory performance. A logistic regression model using heart SUV and heart STR achieved an AUC of 0.86 [95% CI: 0.72–0.96] with a sensitivity of 60%, a specificity of 93% and a false-positive rate of 0.06.
A random forest model incorporating all 10×10 residual correlations and SUV values from the seven most discriminative organs (heart, spleen, lungs, kidneys, brain, liver, and skull/spine) achieved an AUC of 0.80 [95% CI: 0.64–0.95], a sensitivity of 67%, a specificity of 85% and a false positive rate of 0.15.
Conclusions
This study demonstrates that Long COVID is associated with distinct alterations and systemic connectivity changes detectable with 18F-DPA-714 total body PET. The heart was found to be a key organ showing both abnormal SUV and altered network strength, with additional contributions from kidneys and lungs.
Combining SUV and connectivity features provides a promising approach to reveal abnormal inter-organ interactions in Long COVID and supports the potential clinical utility of network-based PET analysis.
Web | DOI | Journal of Nuclear Medicine | Abstract