2.1.

Subject Eligibility and Enrollment

The study was conducted in compliance with an institutionally approved review board protocol at Boston University (protocol number 4698). Informed consent was obtained from all participants prior to the study. Fifteen healthy subjects (eight males and seven females, age $27\pm 4$) were recruited. Each subject participated twice in the study, once after a low-fat meal and once after a high-fat meal. The order in which subjects participated in high-fat or low-fat meal study was randomized. Subjects who were $<18$ years old or who had a prior history of hypotension (low blood pressure), low blood sugar, dizziness, fainting, and type 1 or 2 diabetes were excluded from the study. Participants were asked to fill out a questionnaire containing age, weight, height, race and ethnicity, and skin tone information prior to the start of the study.

2.2.

Postprandial Experiment Procedure

The study procedure is shown in Fig. 1. Subjects who agreed to participate in the study were asked to complete 10 h of overnight fasting. For the high-fat meal study, subjects were provided a meal the next morning that contained $\sim 60\mathrm{g}$ of fat and 1400 KCal. Subjects had the option to choose between breakfast sandwiches from Dunkin Donuts (bacon egg and cheese or sausage egg and cheese on croissant) or a big breakfast option from McDonald’s combined with two protein shakes. For the low-fat meal study, subjects were provided oatmeal (2 g of fat, 110 KCal) in the morning of the study. The nutrient content of the meals is shown in Table S1 in the Supplementary Material. Subjects were monitored following the measurement procedure at the baseline (before consuming the meal in the morning) and every hour for 5 h after the meal.

Fig. 1

(a) Study timeline. (b) Measurement procedure. (c) Schematic diagram of three-LED SFDI system, SFDI data acquisition, and processing flowchart. The top row shows the raw images captured at the spatial frequency of DC and $0.1{\mathrm{mm}}^{-1}$ from the dorsal surface of a subject’s hand at 880 nm. The captured images are demodulated and calibrated, and the diffuse reflectance data is converted into absorption and reduced scattering coefficients using a lookup table. The bottom row shows the ${\mu }_a$ and ${\mu }_s^{\prime }$ maps.

During each SFDI measurement, participants were instructed to sit in a chair and place their right hand into a silicone hand holder positioned under the SFDI system. SFDI measurements were conducted on the dorsal surface of the participants’ hands while they were asked to keep their hands still in the holder. The dorsal hand surface was selected for SFDI measurements due to the visibility of superficial blood vessels. In addition, this area offers convenience for repeated measurements and minimal discomfort for subjects. SFDI patterns were projected at spatial frequency pair of DC and $0.1{\mathrm{mm}}^{-1}$ at three wavelengths of 730, 880, and 1100 nm. Glucose and lipid profiles were then assessed using an in-lab lipid analyzer (Alere Cholestech LDX Analyzer, Waltham, Massachusetts, United States). A fingerstick procedure was performed on the participant’s finger to collect several small drops of blood using a lancet within 10 s. The blood sample was then placed in a cassette and inserted into the Alere system to obtain the full blood profile and glucose concentration.

After finishing the procedure in the fasting state, the subjects were asked to eat either a low-fat (control group) or a high-fat meal (known to induce increased blood lipids). The measurement procedure was then repeated every 1 h for a total of 5 h after the meal. In addition, blood pressure, heart rate, and room temperature were recorded for a subset of subjects (10 high fat and 10 low fat) at every time point. Blood pressure and heart rate were measured using an automated cuff (BPM Connect, Withings). Subjects were free to leave the lab between measurements. Participants were told not to eat for the duration of the study, but they were allowed to drink water. Room temperature data were not recorded for subject 12 in the low-fat meal category, and blood profile data for subject 8 in the same category were missing at the 4-h post-meal timepoint due to complications with the fingerstick blood collection procedure.

2.3.

SFDI Acquisition and Data Analysis

Details regarding the SFDI system and data processing have been described elsewhere.¹³ Briefly, SFDI is a label-free non-contact diffuse optical imaging modality that provides measurements of optical properties (absorption and reduced scattering coefficients) of a biological sample from a large field of view on a pixel-by-pixel basis. In this study, a custom-built LED-based SFDI system was used to perform all the SFDI measurements.¹⁴ The system utilizes two LEDs in the near-infrared wavelength band (730 and 880 nm) and one LED in the shortwave infrared wavelength band (1100 nm) as the illumination source. Both planar and sinusoidal spatial patterns of light at $0.1{\mathrm{mm}}^{-1}$ were projected on a sample using a digital micro-mirror device (DMD), and the remitted light was captured by a germanium CMOS camera (TriWave, Infrared Laboratories, Inc., Peabody, Massachusetts, United States). A three-phase illumination and demodulation technique was used to extract the tissue response at the given spatial frequencies. The field of view (FOV) was $7\times 11\mathrm{cm}$. The exposure time for all three wavelengths was set to 100.925 ms.

Intralipid (Baxter, Deerfield, Illinois, United States) with 10% lipid concentration was used as a calibration phantom with known optical properties to remove the instrument response. The scattering properties of intralipid at a 10% lipid concentration were adapted from literature,¹⁵ and the absorption properties were calculated using Beer’s law for the wavelengths of interest.^16,17 Optical absorption (${\mu }_a$) and reduced scattering coefficients (${\mu }_s^{\prime }$) were estimated at each illumination wavelength from a lookup table generated by Monte Carlo simulations with calibrated diffuse reflectance maps at two spatial frequencies as inputs. In addition, the extracted ${\mu }_a$ values at all three wavelengths were used to extract hemoglobin information, including the concentration of oxyhemoglobin (${\mathrm{HbO}}_2$) deoxyhemoglobin (Hb), total hemoglobin (THb), and hemoglobin oxygen saturation (${\mathrm{StO}}_2$) using Beer’s law. A $2\mathrm{cm}\times 2\mathrm{cm}$ ROI was selected in the center of the hand and the average ${\mu }_a$ and ${\mu }_s^{\prime }$ at the three measurement wavelengths, and the concentrations of ${\mathrm{HbO}}_2$, Hb, THb, ${\mathrm{StO}}_2$, and the a and b scattering parameters were calculated inside the ROI at each time point. The a and b parameters were determined from a power law fit to the three measured ${\mu }_s^{\prime }$ values. In addition, vascular and microvascular areas were segmented using MATLAB fibermetric function and thresholding technique to compare the changes that happen over different anatomic regions.^14,18 We specifically assessed the vascular size of large superficial vessels at each timepoint post-meal, applying a consistent threshold value for segmentation across all timepoints. We also examined the full width at half maximum (FWHM) across prominent superficial vessels to observe any postprandial alterations. A more detailed description of the segmentation process and FWHM calculation is described in the Supplementary Material.

In addition to the absolute optical properties and hemoglobin values, a variety of composite metrics of optical properties were also calculated. All SFDI features were normalized to the value at the baseline to compare the relative changes over time as well as comparing the two categories (low fat versus high fat). Data exhibiting motion artifacts were excluded from the processing analysis to ensure the integrity and accuracy of the results; eight out of 180 (4.44%) optical measurement sessions have excluded datapoints for this reason. The exclusion of the baseline measurement for subject 12, due to motion artifacts after a high-fat meal, necessitated the removal of all timepoints for subject 12 as longitudinal data were normalized to the baseline measurement. Consequently, to maintain the matched low-fat and high-fat dataset, we also excluded the low-fat meal data for subject 12. All data were processed using MATLAB (R2021b, The Mathworks Inc., Natick, Massachusetts, United States) or Python.

2.4.

Statistical Analysis

Descriptive data are presented as mean ± standard error (SE). Statistical analysis was conducted using the “anovan” function in MATLAB to compare the SFDI-derived parameters between two groups (high-fat meal and low-fat meal) over six timepoints and between timepoints at each group. The analysis examined the interaction effect between the group and time factors. Post hoc analysis was performed using the “multcompare” function to determine pairwise differences between the groups at each time point. A $p\text{-}\text{value}<0.05$ was considered statistically significant.

2.5.

Prediction Model

A prediction model was trained to output estimated triglyceride concentrations in the postprandial state using SFDI metrics as inputs. To achieve this, 16 normalized SFDI metrics were used as inputs and trained against triglyceride concentrations relative to their baseline values. A five-fold cross-validation approach was utilized in which the dataset was divided into five equal parts. During each iteration, we trained an XGBoost model on the training dataset and then used it to predict the normalized triglyceride concentrations on the test data. These predictions were then scaled to the original mg/dl scale using their corresponding baseline triglyceride concentrations. This process was repeated five times to ensure the robustness and reliability of our model. We then evaluated our model based on the accuracy of the predictions.

3.1.

Postprandial Changes After Low-Fat and High-Fat Meals

The effect of high-fat and low-fat meals on blood triglyceride, glucose, systolic pressure, diastolic pressure, and heart rate over time is shown in Fig. 2. All results are shown as relative changes from the baseline measurement of the specific parameter. A substantial increase in blood triglycerides occurred at all timepoints after the high-fat meal; by contrast, little change in triglycerides was evident after the low-fat meal across all timepoints. For the high-fat meal subjects, the triglyceride concentrations typically reached their peak 3 h after the meal consumption. These different trends led to statistical differences between high- and low-fat meal cohorts at every hour after the meal ($p<0.001$ for the first hour, $p<0.0001$ for 2 to 5 h). Glucose increased 1 h after the meal for both groups, with significant differences between the groups observed at hours 2 to 4 ($p<0.01$). No significant changes in blood cholesterol occurred following the meals. Systolic blood pressure increased after the high-fat meal, and a significant difference between high-fat and low-fat meals occurred 2 h after the meals ($p<0.05$). Diastolic blood pressure did not show any significant trend with time, but the heart rate increased after a high-fat and decreased after the low-fat meal with a significant difference between the two groups occurring 2 h after the meals ($p<0.05$). Within the cohort (high-fat or low-fat), changes from the baseline were not significant for either systolic or diastolic blood pressure or heart rate. Room temperature was recorded as temperature may affect blood flow in superficial tissue. Room temperature was not significantly different between low- and high-fat cohorts at any time, although there was a significant difference in room temperature at 5 h compared with baseline within the low-fat meal cohort, in which the room temperature increased $2.7\pm 2.0°\mathrm{F}$.

Fig. 2

Relative changes in (a) triglyceride concentration, (b) glucose concentration, (c) systolic blood pressure, (d) diastolic blood pressure, and (e) heart rate compared with the fasting state (time point = 0) during the 5 h postprandial measurements after high-fat (purple) and low-fat (green) meal. Data shows the average values and the standard error as a shaded area. Asterisks show the level of significant differences between the two meal groups. $p<0.0001$ is shown with ****, $p<0.001$ is shown with ***, $p<0.01$ is shown with **, and $p<0.05$ is shown with *.

The temporal changes and trends in SFDI-derived parameters over the 6-h experiment are shown in Fig. 3, including the ${\mathrm{HbO}}_2$ map of a representative subject at baseline and postprandial states after both meals [Fig. 3(a)]. The maps indicate an increase in ${\mathrm{HbO}}_2$ concentration in the vascular and microvascular regions of the subject after the high-fat meal. For this subject, using the $2\mathrm{cm}\times 2\mathrm{cm}$ ROI centered on the hand, the average ${\mathrm{HbO}}_2$ concentration increased from $43.2\mu \mathrm{M}$ at baseline to $94.6\mu \mathrm{M}$ at 3 h after the high-fat meal. Conversely, after the low-fat meal, there was a decline in ${\mathrm{HbO}}_2$ levels. The average ${\mathrm{HbO}}_2$ concentration changed from $54.8\mu \mathrm{M}$ at baseline to $51.5\mu \mathrm{M}$ 3 h after the low-fat meal.

Fig. 3

(a) ${\mathrm{HbO}}_2$ concentration maps at baseline and postprandial states for representative subjects after high-fat and low-fat meals showing an increase in ${\mathrm{HbO}}_2$ after a high-fat meal and a decrease in ${\mathrm{HbO}}_2$ after a low-fat meal. (b) Relative changes in ${\mu }_a$ at 880 nm, (c) $\frac{{\mu }_a(880\mathrm{nm})}{{\mu }_a(730\mathrm{nm})}$, (d) $\frac{{\mu }_a(1100\mathrm{nm})}{{\mu }_a(730\mathrm{nm})}$, (e) ${\mathrm{HbO}}_2$, and (f) ${\mathrm{StO}}_2$ compared with the fasting state (time point = 0) during the 5 h postprandial measurements after high-fat (purple) and low-fat (green) meals. Asterisks show the level of significant differences between the two meal groups. $p<0.0001$ is shown with ****, $p<0.001$ is shown with ***, $p<0.01$ is shown with **, and $p<0.05$ is shown with *.

The longitudinal relative changes of SFDI metrics that changed differentially after the low- and high-fat meals are shown in Figs. 3(b)–3(f). Interestingly, SFDI metrics peaked 3 or 4 h after the high-fat meal, which largely matches the timing when conventionally measured triglyceride concentrations also reached their peak (Fig. 2). In several cases, there were statistically significant differences in SFDI metrics between high-fat and low-fat meals at multiple timepoints after the meal.

A separate examination of large vessel and microvascular regions following segmentation showed no distinct variations in trends between these areas. In addition, a spatial analysis of vascular size and FWHM did not reveal any significant trends or differences in response to meals with varying fat content. The detailed FWHM results, alongside an illustrative segmentation map, are presented in Fig. S1 in the Supplementary Material.

The notable correlations between blood triglyceride levels, blood pressure, and heart rate and SFDI-based metrics are shown in Fig. 4. The correlations for normalized parameters relative to baseline are provided. Overall, nine optical metrics measured with SFDI were correlated with blood triglyceride concentration with $p\text{-}\text{value}<0.05$. The highest correlations were found between triglycerides and $\frac{{\mu }_a(880\mathrm{nm})}{{\mu }_a(730\mathrm{nm})}$ as well as the heart rate and $\frac{{\mu }_a(880\mathrm{nm})}{{\mu }_a(730\mathrm{nm})}$, with a Pearson correlation coefficient = 0.38 in both cases, and $p\text{-}\text{value}={10}^{-5}$ and $p\text{-}\text{value}={10}^{-7}$. These results demonstrate that optical measurements measured with SFDI change in concert with triglycerides and hemodynamic parameters in the postprandial state. No scattering-based metrics correlated strongly with gold-standard measures.

Fig. 4

Correlation plots for highly correlated metrics combining all subjects from both high-fat and low-fat groups. All metrics were normalized to their respective baseline values. Black lines show the linear fit and 95% confidence interval. Green data points indicate measurements from the low-fat cohort, and purple from the high-fat cohort.

3.2.

Prediction of Blood Triglyceride Concentration from SFDI

Figure 5(a) shows predicted versus known triglyceride concentrations in the postprandial state. The correlation coefficient between the predicted and measured values was found to be 0.72 (with a $p$-value of ${10}^{-27}$), and the root mean square error (RMSE) of $40\mathrm{mg}/\mathrm{dL}$. We further explored the ability to predict two key postprandial metrics: the area under the curve (AUC) and the peak triglyceride concentration during the longitudinal assessment. These comparisons are represented in Figs. 5(b) and 5(c), with Pearson correlation coefficients of 0.83 and 0.76, respectively. These results demonstrate the potential of optical measurements to predict important postprandial parameters in this subject population.

Fig. 5

Predictive analysis of triglyceride concentrations utilizing optical metrics with the XGBoost model: (a) A plot of the predicted versus the actual triglyceride concentrations for all subjects across all timepoints. (b) A comparison of the predicted and measured area under the curve (AUC) for triglyceride concentration during the postprandial state. (c) The peak triglyceride concentrations as predicted by the model against the observed maximum values post-meal for each participant.