Near-infrared spectroscopic mapping of the human epicardium

Abstract

Epicardial catheter ablation is necessary to address ventricular tachycardia targets located far from the endocardium, but epicardial adipose tissue and coronary blood vessels can complicate ablation. We demonstrate that catheter-based near-infrared spectroscopy (NIRS) can identify these obstacles to guide ablation. Eighteen human ventricles were mapped ex vivo using NIRS catheters with optical source-detector separations (SDSs) of 0.6 and 0.9 mm. A logistic regression model trained from manually labeled spectra achieved mean area under the receiver operating characteristic curve (AUROC) of 0.907 (0.6 mm SDS) and 0.911 (0.9 mm SDS) in binary adipose detection. Novel optical indices for adipose detection were also proposed, achieving AUROCs of 0.881 (0.6 mm SDS) and 0.873 (0.9 mm SDS), while a blood-specific optical index achieved AUROC of 0.859 for vessel detection (0.9 mm SDS). These results suggest that catheter-based NIRS can detect adipose tissue and coronary vessels to improve efficacy and safety of epicardial ablation.

Graphical Abstract

Epicardial catheter ablation is required to address ventricular tachycardia targets far from the endocardium, but epicardial adipose tissue and coronary blood vessels can complicate the procedure. On 18 donor human ventricles, we demonstrate detection of epicardial adipose tissue and coronary vasculature with catheters retrofitted to perform near-infrared spectroscopy. These results suggest that near-infrared spectroscopy could be a useful tool to improve the efficacy and safety of epicardial ablation.

1 |. INTRODUCTION

Ventricular tachycardia (VT) is a leading cause of sudden cardiac death, which accounts for roughly half of all cardiovascular mortality^[1]. In recent decades, VT treatment has evolved to include implantable cardioverter-defibrillators (ICDs), anti-arrhythmic drugs (AADs), and catheter ablation. While ICDs can terminate life-threatening rhythms, they deliver painful and traumatic shocks that cannot prevent future episodes^[2,3]. Only catheter ablation has the potential to be curative, and ablation has demonstrated greater efficacy than AADs in multiple studies^[2,4,5].

Catheter radiofrequency ablation (RFA) aims to destroy a pathogenic region of myocardium, terminating the arrhythmia and preventing recurrence^[6]. Despite advances in electroanatomic mapping and ablation technology, VT remains difficult to treat– acute procedural success rates are around 75%, and VT recurrence occurs within three years for roughly 50% of patients^[7]. One well-established source of ablation failure is the presence of VT substrates far from the endocardial surface (from which most ablations are performed). These cases necessitate ablation from the epicardial surface, using a catheter introduced into the pericardial space^[2,8–10]. While the effectiveness of epicardial ablation has been established by multiple studies, procedural difficulty and complications have limited adoption^[11–13].

Many challenges of epicardial catheter ablation arise from the presence of unique obstacles on the epicardial surface. The human epicardium is covered with a variable layer of adipose tissue, which can reduce the penetration of RFA^[8,14,15]. Further, epicardial adipose tissue can attenuate catheter-measured electrogram voltages, mimicking scar tissue commonly targeted with ablation^[16–18]. In addition, complications can arise with damage to critical structures such as epicardial coronary vessels. Inadvertent puncture of a coronary vessel can result in serious bleeding, and even ablation delivered in close proximity to coronary arteries can induce pathologic stenosis^[19]. As a result, care must be taken to avoid epicardial vessels, and guidelines recommend against ablation within 5 mm of a major vessel^[8].

Various imaging modalities have been applied to these challenges, including computed tomography (CT) and magnetic resonance imaging (MRI), which have both been used to identify epicardial adipose for ventricular ablation^[20–22]. Pre-ablation angiography is also required to localize coronary vessels and avoid coronary complications^[8]. However, these pre-procedural assessments are expensive, require non-trivial registration with electroanatomic mapping systems, and in the case of CT, involves exposure to ionizing radiation and intravenous contrast agents. Consequently, a tool that efficiently localizes adipose tissue and coronary vasculature could streamline and simplify epicardial ablation procedures.

Near-infrared spectroscopy (NIRS) is a diagnostic modality that can assess local tissue composition. When employed in a diffuse reflectance geometry, NIRS involves illuminating a tissue with near-infrared light, collecting multiply-scattered photons that reemerge, and assessing wavelength-dependent reflectance. This diffuse reflectance contains information about the tissue’s light absorption $\left({\mu }_a\right)$ and sce $\left({\mu }_s^{\prime }\right)$, in a way that depends on the source-detector separation (SDS). These optical properties have been shown useful for tissue differentiation across a range of medical applications, including cancer detection^[23–26], fibrosis quantification^[27,28], adipose tissue identification^[29,30], and monitoring of perfusion and oxygenation^[31–33]. When performed from an electrophysiology catheter, NIRS measurements are intrinsically coregistered with simultaneous electroanatomic maps, unlike CT and MRI.

Our group has demonstrated the ability of catheter-based NIRS to localize RFA lesions on the endocardial surface^[34,35]. Most recently, we validated these NIRS-based measurements of ablated tissue in conditions resembling those in vivo, namely submerged in blood and without perpendicular catheter-tissue contact^[36]. This work utilized the known NIRS signature of blood to identify and reject measurements with poor contact, ensuring that remaining measurements captured only underlying tissue composition. This result suggests that NIRS may be capable of localizing epicardial coronary vessels from the spectroscopic signature of blood. In 2020, our group also published a preliminary demonstration of NIRS to identify epicardial adipose tissue on ex vivo human hearts^[37]. However, this study was performed with a 12 French instrument, significantly larger than the catheters used in electrophysiology procedures today. Newer, smaller catheters necessitate a NIRS source-detector separation less than the 2.31 mm used by Singh-Moon et al ^[37], potentially impacting the ability to detect adipose tissue. In this work, we aim to show that NIRS can identify adipose tissue and coronary vessels on the human ventricular epicardium when deployed from a 7 French catheter with < 1 mm source-detector separation. These results, along with the potential for seamless integration with electroanatomic mapping, suggest that catheter-based NIRS can enhance the safety and efficacy of epicardial ablation.

2 |. METHODS

2.1 |. Spectroscopy Catheter and System

The catheter-based spectroscopy system used in this study was largely identical to that in^[36]. Commercially available 7 French, irrigated ablation catheters (DI7TCDLRT, Biosense Webster, Diamond Bar, CA, USA) were retrofitted to perform near-infrared spectroscopy as follows. A small hole was drilled in each catheter’s tip, allowing two multimode optical fibers to be passed through the catheter’s irrigation channel. These fibers (one for illumination, and one for detection) were terminated at the catheter tip, immobilized with epoxy, and polished on their distal end to maximize light transmission. The illumination fiber (AFS50/125/145T, Fiberguide, Fairfax, VA, USA) measured 125 um in diameter (with a 50 um core), while the detection fiber (FG200LEA, Thorlabs, Newton, NJ, USA) measured 220 um (with a 200 um core). These fibers were fitted with SMA905 connectors (Thorlabs, Newton, NJ, USA) on their proximal ends to ensure robust connections with supporting instrumentation. Of note, these modifications do not significantly interfere with the catheter’s ability to deliver ablation, as demonstrated by previous work^[38]. Two NIRS-capable catheters were fabricated and used during this study, all aspects of which were identical other than the tip spacing between the two optical fibers (the source-detector separation). One of these catheters had a source-detector separation of 0.6 mm (the “0.6 mm catheter”) and the other had a source-detector separation of 0.9 mm (the “0.9 mm catheter”). These source-detector separations were chosen to maximize sensitivity to tissue absorption^[39] within the catheter’s usable space. An illustration of the NIRS-enabled catheter is shown in Figure 1.

FIGURE 1.

Illustration of the NIRS-enabled electrophysiology catheter. A) Diagram showing existing catheter hardware along with placement of integrated optical fibers. B) Frontal view of the catheter face, illustrating the exposed optical fiber faces from where light emerges and is collected. The distance between these fibers (the “source-detector separation”) is crucial to NIRS measurements.

A broadband halogen lamp (HL-2000-HP, Ocean Optics Inc., Dunedin, FL, USA) with emission from approximately 500 to 1100 nm was connected to the source fiber to provide illumination. Diffusely reflected light was collected by the detection fiber and sent to a near-infrared spectrometer (C9405CB, Hamamatsu Corporation, Bridgewater, NJ) sensitive to light from 435 to 1145 nm, with approximately 0.7 nm resolution. The sensor of an electromagnetic tracking system (TrakStar, Northern Digital Inc., Waterloo, ON, Canada) was affixed to the tip of the catheter to record its position and angle throughout spectral acquisition. The accuracy (RMS error) of this tracking system was 1.40 mm in position and 0.5 degrees in angle.

2.2 |. Sample Preparation and Acquisition Protocol

In this study, 11 human donor hearts (18 ventricular specimens) obtained from the National Disease Research Interchange (NDRI, Philadelphia, PA, USA) were imaged ex vivo to demonstrate epicardial NIRS mapping. All specimens were deidentified and not considered human subjects (by the Columbia University Institutional Review Board, under 45 CFR 46). Hearts were dissected to separate atria and ventricles, and the ventricles (left and right) were used in this work. The 18 specimens used in this study are listed in Table 1, along with the NIRS source-detection with which they were mapped, the chamber, and the associated donor demographic information. For clarity, specimens have been numbered 1–18, while donors (which do not match up one-to-one with specimens) are lettered A-K. Following dissection, ventricular specimens were flattened, placed in a dedicated chamber with epicardial surface up, and photographed by a scientific camera (DCC3260C, Thorlabs, Newton, NJ, USA) mounted directly above the chamber. For each image, pixel-wise locations were matched with corresponding physical positions on the specimen (recorded by the tracking system) to generate a set of control points. A projective image transformation matrix was then fit to these control points using MATLAB’s fitgeotrans command in order to map tracker-reported positions onto the image during acquisition. The positional accuracy (RMS error) of registration was measured to be 3.1 mm, corresponding to 28 image pixels at the digital resolution of the camera system.

TABLE 1.

Experimental and demographic information for all 18 ventricular specimens (from 11 donors), each of which was mapped with either the 0.6 mm or 0.9 mm source-detector separation (SDS) NIRS catheter.

Human Heart Specimen and Donor Characteristics
Specimen	SDS	Chamber	Donor	Age	Sex	BMI	Relevant Disease History	Cause of Death
1	0.6 mm	LV	A	67	F	28.8	CAD, CHF, CVA, DM, HLD, HTN, PE, COPD	Respiratory Arrest
2	0.6 mm	RV
3	0.6 mm	LV	B	60	M	43.3	CAD, STEMI, GERD, HLD, HTN, OSA	Cardiac Arrest
4	0.6 mm	RV
5	0.6 mm	RV	C	69	M	26.8	AF, CAD, CKD, CVA, HTN	Cardiac Arrest
6	0.6 mm	LV	D	47	F	30.9	CAD, MI, HLD, DM, HTN, GERD	Anoxia
7	0.6 mm	RV
8	0.6 mm	LV	E	65	M	25.1	AF, CHF, HTN, HLD, COPD, DM, OSA	Anoxia
9	0.6 mm	RV
10	0.9 mm	LV	F	70	M	34.7	CHF, COPD, AF, CM, HTN	Cardiac Arrest
11	0.9 mm	RV
12	0.9 mm	RV	G	58	F	16.4	AF, CHF, HTN, CVA	Stroke
13	0.9 mm	LV	H	59	F	28.2	CHF, HTN, DM, COPD	Cardiac Arrest
14	0.9 mm	LV	I	60	M	40.2	CAD, MI, CHF, HLD, HTN, DM	Anoxia
15	0.9 mm	LV	J	61	M	17.9	CAD, MI, CHF, HLD, HTN, DM	Heart Failure
16	0.9 mm	RV
17	0.9 mm	LV	K	54	M	60.3	CHF, HTN, HLD, DM, COPD	Anoxia
18	0.9 mm	RV

Donor age at death ranged from 47 to 70 years, and body mass index (BMI) ranged from 16.4 to 60.3. LV: left ventricle; RV: right ventricle; AF: atrial fibrillation; CAD: coronary artery disease; CHF: congestive heart failure; CKD: chronic kidney disease; CM: cardiomyopathy; COPD: chronic obstructive pulmonary disease; CVA: cerebrovascular accident; DM: diabetes mellitus; GERD: gastroesophageal reflux disease; HLD: hyperlipidemia; HTN: hypertension; MI: myocardial infarction; OSA: obstructive sleep apnea; PE: pulmonary embolism.

Samples were submerged in phosphate-buffered saline (BP399–20, Fisher Scientific, Waltham, MA, USA) and comprehensively mapped with either the 0.6 mm or 0.9 mm NIRS-enabled catheter. At each measurement location, a full spectrum (with 200 ms exposure time) and the catheter tip position were simultaneously recorded. With the previously determined image transformation, the tracking system allowed for live visualization of the catheter position on the image at rates up to 80 Hz (though exposure time limited the rate of NIRS point acquisition to approximately 5 Hz). Ultimately, it was feasible to collect NIRS maps of approximately one thousand points (each containing a reflectance spectrum and recorded position and angle) in approximately 5 minutes. The data acquisition and image registration procedure is depicted in Figure 2. Experimental control, data acquisition, and processing were performed in MATLAB 2023b (The Mathworks Inc., Natick, MA, USA).

FIGURE 2.

Schematic representation of data acquisition and catheter position registration. First, complete human hearts were dissected to isolate ventricular specimens. Next, specimens were placed in a dedicated chamber, imaged with a scientific camera, and a series of control points were collected for image registration. Specimens were then comprehensively mapped with a custom NIRS-enabled catheter, recording both spectra and catheter positions. Finally, collected spectra were processed into predictions and superimposed onto the images of the specimens.

For one left ventricular sample, additional steps were taken to demonstrate detection of coronary vasculature. A small plastic cannula was sutured in place within the lumen of an epicardial coronary vessel. Anticoagulated porcine blood (Lampire Biological Laboratories, Pipersville, PA, USA) was placed in a bath at 37°C (ED v.2, Julabo, Allentown, PA, USA) and allowed to oxygenate with stirring and exposure to room air. A pulsatile pump (Peri-Star Pro, World Precision Instruments, Sarasota, FL, USA) was then used to circulate blood through the cannulated vessel at a flow rate of approximately 240 mL/min. NIRS mapping was then carried out with the 0.9 mm catheter, ensuring that an adequate density of points were collected in the vicinity of the perfused vessel. The choice to use the 0.9 mm (rather than 0.6 mm) catheter for this task was based on the assumption that NIRS contrast for blood arises primarily from absorption, and so would benefit from longer source-detector separation.

2.3 |. Spectral Labeling and Preprocessing

Using custom-designed MATLAB software, sample photographs were fully segmented with manually-drawn polygons, guided by visualization of individual image channels in the hue-saturation-brightness color space. These polygons were used to generate a set of spectra labeled as either “adipose”, “myocardium”, or “unlabeled”, as follows. First, a polygon was drawn to delineate the visible edges of the tissue based on identification of the (black) image background by low brightness. Spectra whose positions lay outside of this boundary were omitted from further analysis. Next, polygons were drawn to identify areas of clear adipose tissue and exposed myocardium (based on image hue), and spectra within these regions were labeled as “adipose” and “myocardium”. An example of this process for one left ventricular specimen (number 13, from donor H) is shown in Figure 3. These regions were deliberately created to encircle only regions of unambiguous adipose and myocardium, excluding areas with very thin (< 1 mm) or scattered adipose tissue. Additionally, care was taken to leave a distance of approximately 60 pixels (roughly twice the registration resolution) between adjacent regions of different classes so that image registration errors would not result in mislabeling (see Figure 3 B). Finally, all spectra whose positions were within the overall tissue boundary, but not within a region of clear adipose or myocardium, were marked as “unlabeled” (shown in gray in Figure 3 C). These spectra were excluded from training but included in visualizations and when generating predictions. For the sample used to demonstrate epicardial vessel detection, an additional polygon was drawn delineating the vessel’s visible location. Spectra collected from within this boundary were labeled as having come from the vessel, excluded from adipose/myocardium classification, and analyzed independently as discussed below. To mitigate bias, all manual segmentation was performed while blinded to NIRS data, and all segmentation choices were reviewed and verified by a second investigator.

FIGURE 3.

The process of manual labeling, shown step-by-step (for specimen 13, from donor H). A) A photograph of the specimen, with the (manually segmented) background removed and notable landmarks labeled. B) Manually drawn polygons delineating regions of thick adipose tissue (yellow) and clearly exposed myocardium (dark blue). C) The registered positions of measured spectra labeled with manually drawn polygons, colored to indicate the assigned label. Spectra not within a polygon, but within the tissue boundary, were marked as “unlabeled”, shown here in gray.

All retained spectra were subjected to the following preprocessing pipeline. First, spectra acquired with poor catheter-tissue contact were identified by low reflectance and excluded. To avoid rejecting spectra from naturally less reflective regions, a (conservative) threshold was set at 50% of the peak raw reflectance from an absorptive, polydimethylsiloxane optical phantom (remeasured each session). The few spectra whose raw reflectance did not exceed this threshold were deemed to have insufficient contact and excluded from further analysis. Remaining raw spectra were then divided (per-wavelength) by a spectrum collected from a spectrally flat diffuse reflectance standard (Spectralon 99%, Labsphere, North Sutton, NH, USA). This step was performed to calibrate out wavelength-dependent artifacts and day-to-day variation in lamp intensity^[36,40,41]. Each spectrum was smoothed with a moving average filter of length 21 samples, then normalized to have unity integrated area^[28], removing (potentially varying) amplitude differences between spectra. Derivative spectra were also estimated using an 11-sample Savitzky-Golay differentiation filter ^[42]. Due to weaker lamp emission at either end of the spectrum (limiting signal-to-noise ratio), only data between 600 and 900 nm was included for analysis and classification. For ease of processing, spectra were resampled at integer wavelengths (in nanometers). The resulting data (preprocessed spectra, each with a categorical label) were then partitioned by catheter source-detector separation into a 0.6 mm and 0.9 mm dataset (each containing nine specimens). The same preprocessing was applied to all spectra, regardless of source-detector separation or classification method.

2.4 |. Spectral Classification

Two distinct classification approaches were used to predict categorical labels from spectral information. In all cases, spectra collected by the 0.6 and 0.9 mm catheters were analyzed separately. This was necessary due to the significantly different spectral shapes produced by the different source-detector separations.

2.4.1 |. Principal Component Analysis and Regression

First, principal component analysis (PCA) was used to identify the linear combinations of wavelengths responsible for the most variation across spectra. In this process, the mean spectrum was first subtracted off to center the distribution of spectral values at each wavelength at zero. Then, PCA was used to identify the principal component spectra (i.e. spectral basis vectors) responsible for the most variation across all spectra (of a given SDS), along with principal component scores (projections of individual spectra along these basis vectors). Component-wise fractions of total variance were used to identify a small number (between 2 and 10) of components to keep, based on the principle of locating the “knee” in explained variance^[43]. Specifically, this was accomplished by searching for an abrupt change in the slope between successive values, and then keeping only preceding components^[44]. Retained component scores were then used as regressors in a logistic regression model (using MATLAB’s glmfit command), estimating a probability of adipose tissue for each spectrum. Training and testing of this model was performed using exhaustive leave-one-out cross validation, withholding a single test specimen at a time from training. All data (excluding blood vessel spectra) were included in the model, and all specimens were used as the test specimen exactly once. Further details about the development and evaluation of this model can be found in the Supporting Information.

With each train-test split, probability estimates were generated for all spectra within the unseen test specimen (including “unlabeled” spectra). These were visualized using color to display probability, both with discrete scatter plots and interpolated continuous maps. Estimated probabilities for spectra with known labels (“adipose” or “myocardium”) were assessed by area under the receiver operating characteristic curve (AUROC). These probabilities were also thresholded into binary predictions (using a probability threshold of 0.5) and assessed with accuracy, sensitivity, specificity, and Matthews Correlation Coefficient (MCC), assuming adipose to be the “true” class. Due to significant variation across specimens, results were also generated after standardizing principal component scores (subtracting the mean, and dividing by the standard deviation) within each specimen. All results are reported as means and standard deviations across all train-test splits, both before and after this per-specimen standardization. Independent two-sided t tests were performed to assess whether there was a difference in these continuous metrics between the 0.6 mm and 0.9 mm catheters, and whether there was improvement after score standardization. A significance level of 0.05 was used for all hypothesis tests.

2.4.2 |. Optical Index Determination and Evaluation

Beyond principal component analysis, we sought to generate predictions using simple, spectrally-derived optical indices using information at only a select few wavelengths. This approach was motivated by the potential to implement these metrics using inexpensive LEDs and a wavelength-insensitive photodetector, rather than a costly lamp and spectrometer. To ensure insensitivity to overall spectral amplitude, these optical indices were chosen to be ratios of reflectance at two individual wavelengths. In order to discern the optimal choices for numerator and denominator wavelengths, a grid search of all possible combinations was performed. For each combination, the collective AUROC (across all nine specimens of a single SDS) in distinguishing adipose from myocardium was used as the figure of merit. After lowpass filtering, nearby spectral values are highly correlated, limiting potential contrast from differences in values at nearby wavelengths. For this reason, combinations of wavelengths within 50 nm of one another were excluded from consideration. For the same reason (and to reduce computational burden), spectral values were tested in 5 nm increments. The combinations of wavelengths (one for each source-detector separation) yielding the highest AUROC values were then selected as optimal ratiometric optical indices (termed “adipose contrast indices”, or ACIs). The technique of per-specimen standardization was also applied to ACI values (though grid searches were performed without it), and overall AUROC values were computed with and without this standardization. Two-sided Z tests were used to assess equality of these AUROCs between source-detector separations, and to assess improvement with standardization, using the method described by Hanley and McNeil^[45] to estimate AUROC standard errors.

To demonstrate epicardial vessel detection, an optical index previously shown capable of detecting both oxygenated and deoxygenated blood was employed. The formula for this metric, the contact optical index ($\mathrm{C}\mathrm{O}\mathrm{I}$), is shown below^[34]:

$$COI=\frac{R\left(764\mathrm{n}\mathrm{m}\right)}{R\left(730\mathrm{n}\mathrm{m}\right)}$$

Because this optical index was originally developed to identify catheter-tissue contact (i.e. a lack of blood), we hypothesized that the perfused vessel would be identified by low $\mathrm{C}\mathrm{O}\mathrm{I}$ values. These spectrally-derived $\mathrm{C}\mathrm{O}\mathrm{I}$ values were used with manual vessel labels to evaluate the performance of this metric via AUROC (assuming lower values to indicate vessel).

3 |. RESULTS

3.1 |. Spectral Labeling and Preprocessing

Initially, 25,600 spectra were collected across all 18 specimens. Of these, 289 spectra (1.1%, or 16.0 per specimen) were deemed to have poor contact and excluded, leaving 25,311 spectra (1,406.2 per specimen). Based on the total two-dimensional surface area mapped, this represented a spatial sampling density of 11.6 spectra per cm². Of these, 8,249 were acquired with the 0.6 mm catheter, and 17,062 were acquired with the 0.9 mm catheter. After labeling, the class break-down (for non-vessel spectra) was as follows: 14,326 (56.6%) adipose, 4,201 (16.6%) myocardium, and 6,784 (26.8%) unlabeled. Considering only those spectra with known labels (as occurred in classifier training and testing), the breakdown was 77.3% adipose and 22.7% myocardium. These percentages were relatively consistent across the 0.6 mm and 0.9 mm datasets. The mean spectra (and derivative spectra) for adipose and myocardium, for both source-detector separations, are shown in Figure 4. These results illustrate the significantly different spectral morphologies resulting from the different source-detector separations.

FIGURE 4.

Mean spectra (A and C) and derivative spectra (B and D) for adipose and myocardium, for the two source-detector separation (SDS) values of 0.6 mm and 0.9 mm, respectively. Shaded blue and yellow regions show one standard error above and below the mean, calculated from the nine mean spectra of each source-detector separation. Gray vertical bars in A and C indicate the numerator and denominator wavelengths selected for adipose contrast indices. Derivative spectra were estimated using an 11-point Savitzky-Golay differentiation filter.

3.2 |. Spectral Classification

3.2.1 |. Principal Component Analysis and Regression

As previously discussed, principal component analysis was applied independently to non-vessel spectra from the 0.6 mm and 0.9 mm datasets. Initially, this process generated 301 principal components (because original spectra were resampled to contain values at this many wavelengths). Based on successive changes in component-wise explained variance, the three leading principal components were retained for the 0.6 mm data, and the four leading components were retained for the 0.9 mm data. Figure 5 shows these retained principal component spectra, along with “scree” plots showing fractional explained variance versus component number. Figures 6 A–6 D contain box and whisker plots of leading principal component scores (projections of spectra onto the first component) within each specimen. Scores for specimens mapped with the 0.6 mm source-detector separation are shown (before and after standardization) in Figures 6 A and 6 B, with analogous information for 0.9 mm in Figures 6 C and 6 D. Retained principal component scores were then compiled into feature vectors (three-element vectors for the 0.6 mm data, and four-element vectors for the 0.9 mm data) for logistic regression model training and testing. Quantitative model performance metrics (AUROC, accuracy, sensitivity, specificity, and MCC) are shown in Table 2 for both source-detector separations, before and after per-specimen score standardization. The full results of t tests comparing these metrics are shown in Supplementary Table S1. Of these, the only statistically significant result was the increase in MCC with standardization, for the 0.9 mm SDS (p = 0.008).

FIGURE 5.

Retained principal component spectra for the 0.6 mm (A) and 0.9 mm (B) source-detector separations. Inset plots show the fractions of total variance explained by the first six principal components, and red dotted lines indicate the cutoffs used to determine the number of components to retain. These cutoffs were chosen to lie just before a “knee” in each the plot, and resulted in three and four retained components for the 0.6 mm and 0.9 mm data, respectively.

FIGURE 6.

Individual-specimen box and whisker plots of leading principal component scores. A, B) Distributions of leading principal component (PC₁) scores before standardization, for 0.6 mm and 0.9 mm source-detector separations, respectively. C, D) Distributions of PC₁ scores for 0.6 mm and 0.9 mm, after per-specimen standardization. Letters (“L” or “R”) after specimen numbers indicate chamber (left or right ventricle). Boxes extend from the first to third quartile, with the midline indicating the median. Whiskers extend to the minimum and maximum values.

TABLE 2.

Results for logistic regression classification of adipose tissue and myocardium, using retained principal component scores as regressors.

Mean (and Standard Deviation) Logistic Regression Classification Results
Source-Detector Separation	0.6 mm (n = 9)		0.9 mm (n = 9)
Scores Standardized?	NO	YES	NO	YES
AUROC	0.886 ± 0.090	0.907 ± 0.066	0.876 ± 0.100	0.911 ± 0.034
Accuracy (Prob. ≥ 0.5)	0.844 ± 0.116	0.854 ± 0.133	0.750 ± 0.129	0.826 ± 0.087
Sensitivity (Prob. ≥ 0.5)	0.957 ± 0.038	0.955 ± 0.035	0.902 ± 0.187	0.931 ± 0.038
Specificity (Prob. ≥ 0.5)	0.474 ± 0.357	0.614 ± 0.293	0.261 ± 0.411	0.588 ± 0.250
MCC (Prob. ≥ 0.5)	0.459 ± 0.294	0.564 ± 0.263	0.288 ± 0.179	0.515 ± 0.135

Each result represents the mean (± standard deviation) of nine train/test splits. Other than AUROC, all metrics were calculated after thresholding estimated probabilities with a cutoff value of 0.5. AUROC: area under the receiver operating characteristic curve; MCC: Matthews correlation coefficient.

In addition to these quantitative metrics, model performance was assessed qualitatively through visualization of predicted probabilities (encoded by color) superimposed onto specimen images. An example is shown in Figure 7 (for specimen 13), with two different visualization methods, alongside the manual labels (Figure 7 A). In Figure 7 B, the principal component score-based logistic regression model is used to generate adipose probabilities for all spectra, including those originally unlabeled. In Figure 7 C, (linear) interpolation has been used to estimate probabilities everywhere within the boundary created by the scattered points.

FIGURE 7.

Visualizations of adipose/myocardium predictions superimposed on an image of a left ventricular specimen (number 13, from donor H). A) Manually labeled regions of thick adipose tissue (yellow) and clearly exposed myocardium (dark blue). B) Color-encoded adipose probability predictions (from principal component-based logistic regression) shown as discrete points at the positions of individual spectra, as recorded by the tracking system. C) The interpolated probability prediction map created using (linear) interpolation between the individual data points in B. D) Color-encoded adipose contrast index ($AC{I}_{0.9}$) values shown as discrete points at tracked positions. E) The interpolated map of adipose contrast index from data points in D.

3.2.2 |. Optical Index Determination and Evaluation

As described, a grid search was performed to evaluate the ability of all ratiometric combinations of spectral values to distinguish between adipose tissue and exposed myocardium. Excluding pairs of spectral values within 50 nm of one another, 2652 candidate optical indices remained. In fact, only half of these were unique, because inverting values simply reverses ordering and changes AUROC to one minus its original value. The best-performing optical indices (adipose contrast indices) for the 0.6 mm and 0.9 mm source-detector separations were:

$$AC{I}_{0.6}=\frac{R\left(870\mathrm{n}\mathrm{m}\right)}{R\left(785\mathrm{n}\mathrm{m}\right)}$$

$$AC{I}_{0.9}=\frac{R\left(610\mathrm{n}\mathrm{m}\right)}{R\left(755\mathrm{n}\mathrm{m}\right)}$$

Full maps of candidate optical indices are shown in Supplementary Figure S1, indicating where on these maps optimal chosen indices lie. Box and whisker plots of these adipose contrast index values for individual specimens are shown in Supplementary Figure S2. Despite variation in ACI values between specimens, adipose tissue showed consistently higher values than myocardium. Across all specimens mapped with the 0.6 mm catheter (numbered 1–9), the AUROC of the $AC{I}_{0.6}$ was 0.887, and 0.881 with standardization. This difference was not statistically significant (p = 0.337). For those mapped with the 0.9 mm catheter (numbered 10–18), the AUROC of the $AC{I}_{0.9}$ was 0.822. Standardization of $AC{I}_{0.9}$ values yielded an AUROC of 0.873, a statistically significant improvement (p < 0.001). Without standardization, the $AC{I}_{0.6}$ achieved higher AUROC than $AC{I}_{0.9}$ (p < 0.001), but this statistically significant difference disappeared after standardization (p = 0.121). For one example specimen (number 13), scattered and interpolated visualizations of the $AC{I}_{0.9}$ are displayed in Figures 7 D and 7 E, in an analogous fashion to model-predicted probabilities.

The perfused specimen (specimen number 15, from donor J) contained a total of 2,571 spectra. Of these, 187 (7.3%) were manually labeled as having been taken from the vessel. Figure 8 A depicts a portion of this specimen, with the perfused vessel highlighted in red. The receiver operating characteristic curve computed from all contact optical index values within this specimen had AUROC of 0.859. Figure 8 B and 8 C depict color-coded contact optical index values as individual points, and an interpolated map, respectively.

FIGURE 8.

A) Image of the perfused ventricle (specimen number 15, from donor J), showing the perfused epicardial vessel in red. B) Discrete NIRS acquisition points from the perfused specimen, color-coded by contact optical index value. Point positions are the result of applying the fitted image registration transform to raw positional information. C) Continuous visualization of contact optical index value, computed by (linearly) interpolating the scattered data points shown in B, showing the detected location of the perfused vessel.

4 |. DISCUSSION

This work aimed to show the ability of catheter-based near-infrared spectroscopy to identify clinically relevant features of the human epicardial surface ex vivo. By tracking the position of the catheter and registering recorded positions to photographs of each specimen, individual spectra were labeled as “adipose tissue,” “exposed myocardium,” or “coronary vessel.” These comprehensive labels allowed for the development and validation of two distinct approaches to spectral classification: principal component analysis-driven logistic regression, and ratiometric optical index calculation. Both of these methods performed well in distinguishing epicardial adipose tissue from myocardium, as evidenced by AUROC values of 0.886 and 0.887 (0.6 mm source-detector separation), and 0.876 and 0.822 (0.9 mm source-detector separation). Due to variation in optical metrics across mapped hearts, the performance of both classification methods improved slightly with standardization of values within each specimen. These results suggest that NIRS, with 0.6 mm or 0.9 mm source-detector separation, can accurately identify adipose tissue for epicardial ablation. Finally, for the perfused specimen, the previously published^[34] contact optical index identified spectra from the perfused vessel with an AUROC of 0.859, suggesting that coronary artery localization may also be feasible with this system. Compared with competing imaging modalities, catheter-based NIRS is fast, inexpensive, and naturally coregistered with electroanatomic maps.

The results of this study affirm the impact of source-detector separation (SDS) on spectral morphology, a well-established result ^[38,39,46]. For small source-detector separations (< 1 mm) and scattering coefficients typical of cardiac tissues in the near-infrared $\left({\mu }_s^{\prime }<30{\mathrm{c}\mathrm{m}}^{-1}\right)$, reflectance is predicted to increase with tissue scattering $\left({\mu }_s^{\prime }\right)$ ^[38,47]. For both adipose and myocardium, reflectance measured with the 0.6 mm SDS was seen to decrease monotonically with wavelength (in Figures 4 A–4 B). Given the preferential sensitivity to scattering conferred by small SDS, this spectral morphology is likely due to the analogous trend in ${\mu }_s^{\prime }$ that characterizes both Rayleigh and Mie scattering. However, Jacques^[48] reports that this decrease in ${\mu }_s^{\prime }$ is less pronounced for fatty tissue than other soft tissues, accounting for the higher long-wavelength reflectance seen here for adipose (and captured by the $AC{I}_{0.6}$). Conversely, a larger SDS results in the collection of photons having traversed deeper, longer paths in tissue, increasing sensitivity to tissue absorption (relative to tissue scattering)^[49,50]. While 0.9 mm represents only a small increase in SDS from 0.6 mm, an increased sensitivity to absorption is clearly evident in our 0.9 mm spectra (in Figures 4 C–4 D). This is most notable in the downward deflection seen for myocardium at 760 nm, likely a result of the absorption peaks of deoxygenated hemoglobin and myoglobin at this wavelength^[48,51]. Expanding the range of wavelengths analyzed in further studies would likely illustrate additional absorption features, such as hemoglobin/myoglobin absorption between 500 and 600 nm, and the lipid absorption peak at 930 nm^[48,52].

The ability of NIRS to distinguish adipose tissue from myocardium at both source-detector separations suggests that contrast exists from both tissue scattering and absorption, and with smaller source-detector separation than previously shown^[37]. This contrast allowed for effective classification using two data-driven approaches: PCA with logistic regression, and grid-search derived optical indices. The PCA-based classifiers slightly outperformed the adipose contrast indices, a result which is not surprising given that PCA was allowed to make predictions using all wavelengths (whereas each optical index used only two). Still, the relatively high AUROCs achieved by the adipose contrast indices suggest that the number of wavelengths used for spectroscopy could be greatly reduced without much loss in classifier performance, a tradeoff that might be worthwhile for a simple, inexpensive device.

Despite the high specificity of tissue absorption, the longer 0.9 mm source-detector separation did not perform consistently better. One potential reason for this may have been significant compositional variation across specimens. If this variation occurred predominantly in tissue absorption (say, from varying chromophore composition with disease states), its impact would be greater with larger source-detector separation. Indeed, significant variation in disease history was present in the donor population (as seen in Table 1 ), and potentially more so than would be expected in patients receiving epicardial catheter ablation. Many donors here had a history including prior cardiac surgery, a procedure known to affect epicardial composition that is typically also a contraindication to epicardial catheter ablation^[13,53–55]. Hence, the variation in spectrally derived metrics observed here (motivating per-specimen standardization) might be less pronounced in an actual patient population.

These results also suggest the feasibility of localizing coronary vasculature with NIRS, to prevent ablation-related coronary injury. A previously validated contact optical index identified the perfused vessel with high AUROC, facilitated by the significantly different spectral signatures of blood and tissue^[48]. However, given the small sample size and potential differences between benchtop and true physiologic perfusion, in vivo studies are necessary to confirm the ability of NIRS to identify vessels in epicardial ablation procedures.

Whether 0.6 mm or 0.9 mm, the small source-detector separation imposed here by catheter diameter limits light penetration to at most 1–2 mm. As a result, this instrument is best-suited to measurements of tissue features close to the surface, such as localization of epicardial ablation obstacles. For depth-sensitive measurements (such as quantification of adipose thickness or detection of deeply embedded coronary vessels), a larger source-detector separation device would be useful. It may also prove useful to incorporate multiple source-detector separations into a single device and leverage their complementary information for depth discriminatory classification.

5 |. CONCLUSION

In summary, we have performed ex vivo epicardial mapping of 18 human ventricles with a prototype NIRS-enabled catheter. We have demonstrated accurate identification of adipose tissue with a regression model using full spectra, and with simple metrics using only select wavelengths. We have also presented evidence to suggest that NIRS can detect coronary vessels via the optical signature of blood, though further validation is needed. These results suggest that catheter-based NIRS can serve as an important tool to improve the safety and efficacy of epicardial ablation procedures.

DATA AVAILABILITY STATEMENT

The data presented in this article are will available upon publication in Columbia University Academic Commons at https://doi.org/10.7916/bjaf-xf68.