2.1.

Optical RFA Catheter and NIRS System

A commercial open-tip irrigated Celsius thermocool RFA catheter (DI7TCDLRT, Biosense Webster, Irvine, California, United States) was modified to accommodate two multimode optical fibers, for illumination and detection. The illumination fiber (AFS50/125/145T, Fiberguide, Fairfax, Virginia, United States) had a core diameter of $50\mu \mathrm{m}$ and a full diameter (core and cladding) of $125\mu \mathrm{m}$. The detection fiber (FG200LEA, Thorlabs, Newtown, New Jersey, United States) had a core diameter of $200\mu \mathrm{m}$ and a full diameter of $220\mu \mathrm{m}$. The fiber tips were immobilized with epoxy during fabrication such that their separation distance was 0.9 mm, as shown in Fig. 1. This source–detector separation was chosen to maximize optical path length and sensitivity to chromophore absorption³² without interfering with existing internal hardware of the catheter, with outer diameter of 2.33 mm (7 French). The source fiber was connected to a broadband light source (HL-2000 HP, Ocean Optics Inc, Dunedin, Florida, United States) to illuminate the tissue surface. The emission spectrum of this light source ranged from 500 to 1100 nm, with nominal output power of 4.5 mW. The optical power measured at the catheter tip was 0.61 mW. Diffusely reflected light was collected by the detection fiber, subjected to a longpass filter with cutoff wavelength 600 nm (FELH0600, Thorlabs, Newton, New Jersey, United States), and analyzed by an NIR spectrometer (C9405CB, Hamamatsu, Bridgewater, New Jersey, United States) with a spectral range from 435 to 1145 nm and a resolution of 0.9 nm. The probe of an electromagnetic tracking system (TrakStar, Northern Digital Inc., Waterloo, Ontario, Canada) was mounted to the tip of the catheter in order to monitor the position and angle of the catheter. Our modifications do not compromise or interfere with the ability of the catheter to deliver RF ablation. Hence, used in conjunction with an RF generator (Stockert 70, Biosense Webster, Irvine, California, United States), the catheter is capable of delivering and continuously monitoring ablation (as we have previously demonstrated in Ref. 28). In addition, positional information from the tracking system allows for precise and instantaneous localization of the catheter tip for each acquired spectrum. Using this information, it is straightforward to construct visualizations of spectrally derived metrics with thousands of points and densities comparable to EAM.¹⁰ All experimental measurements were recorded and processed in MATLAB 2015b and 2019a (The Mathworks Inc., Natick, Massachusetts, United States).

Fig. 1

Commercial ablation catheter with integrated optical fibers for NIRS mapping. (a) Photograph of the catheter, showing the intact mapping electrodes as well as the modified tip. (b) Photomicrograph of the catheter tip illustrating the positions of the two multimode optical fibers.

2.2.

Experimental Samples and RFA Protocol

Experiments were implemented on 22 healthy pig hearts (Green Village Packing Co., Green Village, New Jersey, United States). Each left atrium (including the atrial septum) was separated from ventricles and right atrium and sliced open to lie flat. The tissue was placed (with its endocardial surface up) on an electrosurgical grounding pad in a bath of phosphate buffered saline (PBS) at 37°C. First, irrigated lesions were created on the endocardial surface, with 25 to 45 W of RF power for 15 to 50 s, whereas an irrigation pump (CoolFlow, Biosense Webster, Irvine, California, United States) maintained a flow (of PBS) of $17\mathrm{ml}/\min$ (based on clinical ablation settings) to the catheter tip. The variation in ablation parameters was intended to compensate for varying tissue thickness and avoid complications of overtreatment, such as steam-pops. Between 6 and 9 lesions were generated on each left atrium for a total of 173 lesions. Following lesion creation, left atria were placed in a separate chamber and pinned onto a corkboard for lesion characterization. For lesion characterization, all atria were mapped while submerged in both PBS or fresh porcine blood, to validate that NIRS-derived metrics retain sensitivity in blood. All imaging was completed within two days postmortem. The three main steps of this experimental protocol are depicted in Fig. 2.

Fig. 2

Experimental flow diagram of left atrial NIRS mapping. First, each left atrium was dissected away from the heart and flattened. Then irrigated lesions were generated in a line adjacent to a pulmonary vein. Finally, spectra were acquired from across the endocardium while the tracking system simultaneously recorded the catheter’s position.

Prior to mapping, a reference image of each dissected atrium was captured. For a handful of pixel locations in this reference image, the corresponding positions in physical space were recorded using the tracking system. These control points were used to fit a projective image transformation between physical space and image pixels, using MATLAB’s “fitgeotrans” command. Then the sample was submerged in PBS or blood, and NIR reflectance spectra were obtained for 500 to 2000 individual points, in a continuous acquisition, with a sampling rate of 2 Hz. At each point, the tracking system recorded the position ($X$, $Y$, and $Z$) and orientation (pitch, yaw, and roll) of the catheter tip. For each tracker-derived catheter position, the aforementioned image transformation was used to obtain the corresponding location (in pixels) on the reference image. In this way, the system provided real-time, image-registered visualization of the catheter tip location. The precise number of spectra obtained from each sample was chosen to fully cover the left atrial endocardium, excluding the left atrial appendage. Following NIRS mapping, atria were stained with 1% 2,3,5-triphenyl-2H-tetrazolium chloride (TTC) and further dissected to verify the extent of RFA lesions.

2.3.

Spectral Processing and Optical Index Calculation

NIRS measurements were calibrated and absolute reflectance spectra were converted into relative reflectance spectra (denoted $R_{\mathrm{rel}}(\lambda )$) as follows. Raw spectra were first divided by reflectance from a diffuse (99%) reflectance standard (Spectralon, LabSphere, North Sutton, New Hampshire, United States) to compensate for the spectral shape of the light source and any other wavelength-dependent effects of the system.³³ Next, spectra were scaled by the measured reflectance from a (polydimethylsiloxane) phantom at a wavelength with known optical properties.^34,35 The resulting relative spectra were truncated to exclude wavelengths below 600 and above 1000 nm, lowpass filtered with cutoff frequency $0.05\pi$ rad/sample, and renormalized to their values at 600 nm. This second, per-spectrum normalization was performed to address any remaining variation in spectrum amplitude not corrected by the aforementioned calibration scheme. The choice of normalization wavelength (600 nm) was mainly chosen for simple visualization and was also used in a previous publication.²⁸ Derivative spectra are denoted $R_{\mathrm{rel}}^{\prime }(\lambda )$.

A previously published contact optical index (COI), calculated as shown in Eq. (1), was used to filter out points with poor catheter-tissue contact.²⁸ The threshold value for the COI was evaluated based on the previously published mean COI values for contact and noncontact.²⁸ Within this study, a threshold value of 0.93 was chosen, resulting in 18.5% of the spectra being rejected due to poor contact. After contact-based thresholding, 39,516 preprocessed spectra remained (22,821 from PBS and 16,695 from blood) across 22 atria. They were manually labeled (as either nonlesion or lesion) by comparing their recorded positions with the known positions of lesions. Then optical indices were computed for each acquired spectrum. Many of these optical indices have been previously demonstrated effective in differentiating ablated from unablated tissue, albeit in PBS, with perpendicular contact. The formulas for these optical indices are provided in Eqs. (2)–(5), (LOI, lesion optical index and SOI, substrate optical index). Note that the optical index denoted here as ${\mathrm{LOI}}_0$ was previously published as simply “LOI” and has been renamed here for clarity. One new ratiometric optical index [${\mathrm{SOI}}_4$, shown in Eq. (6)] was devised via a systematic search across all possible numerator–denominator combinations of integer wavelengths:

For each optical index, a two-tailed $t$-test was performed against the null hypothesis of equal means for nonlesion and lesion tissue (pooling data across all hearts). Receiver operating characteristic (ROC) curves were generated and the corresponding areas under the curve (AUCs) were computed to assess the classifying capability of each optical index. Next, for each of the 44 maps (22 in PBS and 22 in blood), a per-heart mean optical index value was calculated for both classes (i.e., nonlesion and lesion). This yielded, for each optical index, 88 values: 22 for nonlesion tissue in PBS, 22 for lesion tissue in PBS, 22 for nonlesion tissue in blood, and 22 for lesion tissue in blood. Two-sided $t$-tests were performed to compare the mean values between PBS and blood for nonlesion and lesion. Figure 3 illustrates this statistical analysis for two optical indices, ${\mathrm{LOI}}_0$ and ${\mathrm{SOI}}_1$. All signal processing and statistical analysis were performed in MATLAB 2019a and GraphPad Prism 9.4.1 (GraphPad Software, United States).

Fig. 3

Spectral morphologies and optical index analysis for two key optical indices. (a) Mean (dotted line) and 95% confidence interval (shaded region) relative reflectance spectra for nonlesion and lesion tissue. (b) Relative reflectance first-order derivative spectra for nonlesion and lesion tissue. (c) Distributions and $t$ test results for ${\mathrm{LOI}}_0$ between nonlesion and lesion tissue. (d) Hypothetical ROC curve and AUC, using ${\mathrm{LOI}}_0$ as an independent classifier. (e) Mean per-heart ${\mathrm{LOI}}_0$ values (with 95% confidence intervals) for nonlesion and lesion tissue, in PBS and blood; (f) Distributions and $t$ test results for ${\mathrm{SOI}}_1$ between nonlesion and lesion tissue. (g) Hypothetical ROC curve and AUC, using ${\mathrm{SOI}}_1$ as an independent classifier. (h) Mean per-heart ${\mathrm{SOI}}_1$ values (with 95% confidence intervals) for nonlesion and lesion tissue, in PBS and blood (****$p<0.0001$, ns: $p>0.05$).

A separate, dedicated experiment on two left atria was conducted to evaluate the sensitivity of NIRS measurements to catheter incidence angle. With the location of the catheter tip held constant, the angle of the catheter was systematically varied from normal incidence (0 deg) to $\sim 70\mathrm{deg}$, as illustrated by Fig. 4(a). Afterward, COI-based thresholding (as described above) was used to exclude spectra obtained with poor contact. This analysis was conducted separately on untreated and ablated left atrial tissue, with samples submerged in blood. Spectra were processed into optical indices as detailed above. The relationship between contact angle and each optical index value could then be evaluated for nonlesion and lesion tissue.

Fig. 4

Contact angle sensitivity assessment. (a) Depiction of dedicated experiments conducted to acquire data at varying catheter contact angles, from a single location. (b) ${\mathrm{LOI}}_0$ data acquired at varying contact angles on lesion and nonlesion tissue. (c) ${\mathrm{SOI}}_1$ data acquired at varying contact angles on lesion and nonlesion tissue. In panels (b) and (c), lines and shaded regions illustrate sliding kernel-weighted estimates of mean and standard deviation.

2.4.

Classification Algorithm

The classification dataset was constructed by calculating a feature vector of optical indices from each individual spectrum,²⁸ including both PBS and blood data. Correlations between features were assessed using the Pearson correlation coefficient, and five representative uncorrelated features were selected to be used in the classification model for the lesion and nonlesion tissue. The dataset was divided into training (32,392 spectra from 18 hearts) and testing (7124 spectra from 4 hearts) groups. The training set contained 28,482 spectra (87.9%) labeled as nonlesion and 3910 (12.1%) labeled as lesion, and the testing set contained 6061 (85.1%) labeled as nonlesion and 1063 (14.9%) labeled as lesion. Five different machine learning classification algorithms were trained and their results compared on the basis of ROC AUC. These models were logistic regression, $k$-nearest neighbors (kNN), extreme gradient boosting (XGB), a support vector machine (SVM), and a random forest (RF). For each model, five-fold cross validation was used within the training set, and the final trained model parameters were obtained by averaging the five results. Class weights were tuned to alleviate the impact of class imbalance. Model training and testing were implemented in MATLAB and RStudio (Posit, Boston, Massachusetts, United States).

2.5.

Blood Mapping Robustness

To assess the equivalence of mapping in PBS and blood, an additional analysis was performed to directly compare model-predicted lesion probabilities for hearts within the test set. First, in order to apply traditional image-processing techniques, scattered points were (linearly) interpolated onto a uniform grid. This interpolation was applied to both the model-predicted probabilities and the ground truth labels. The scattered points from both PBS and blood data were used to define a common concave boundary, outside of which, interpolated data were excluded from analysis. Interpolated data were subjected to a 2D lowpass filter with cutoff frequency $0.1\pi$ rad/sample.

Interpolated model probability data were compared using three different image comparison metrics: structural similarity (SSIM) index,³⁶ Frobenius normalized inner product, and root-mean-squared (RMS) difference. Only points within the predefined boundary were included in these computations. Maps were compared after thresholding probabilities into binary lesion/nonlesion predictions. Binary prediction maps in PBS and blood were compared using the Dice coefficient.

3.1.

Optical Index Assessment

3.2.

Classification Results

Fifteen optical indices were considered as potential features in the classification model [listed in Fig. 5(a)]. ${\mathrm{LOI}}_0$ and ${\mathrm{SOI}}_1$ were preserved for the final model, due to statistical performance illustrated above and in a previous study.²⁸ Three other optical indices were chosen for inclusion based on low correlations with ${\mathrm{LOI}}_0$, ${\mathrm{SOI}}_1$, and each other [highlighted in Fig. 5(a)]. These optical indices were ${\mathrm{LOI}}_8$, ${\mathrm{SOI}}_3$, and ${\mathrm{SOI}}_4$ with individual AUC values ranging from 0.517 to 0.643.

Fig. 5

Assessment of machine learning classifiers. (a) Candidate feature correlation matrix with selected features (i.e., optical indices) highlighted in yellow. (b) ROC curves for tested classifiers. (c) Sensitivity, specificity, and ROC AUC values for tested models (with a probability threshold of 0.5).

All five machine learning models were evaluated on the test set, generating the ROC curves and AUCs shown in Fig. 5(b). All models were effective in the classification of lesion and nonlesion spectra, validating the choice of input features. When thresholded at a probability value of 0.5, all models exhibited high specificity, and the RF classifier yielded the highest sensitivity (81.66%, compared with 48.64%, 48.26%, 57.42%, and 57.86%, for logistic regression, kNN, XGB, and SVM, respectively). The RF model also led to a higher AUC (0.921) than the other tested models [0.914, 0.842, 0.854, and 0.848, shown in Fig. 5(c)].

Based on these results, the trained RF classifier was selected as the final model. An example comparing model predictions to ground truth data for test heart 1 is shown in Figs. 6(a)–6(c). Figure 6(a) shows the reference image with the true lesion locations outlined (intentionally generated alongside the boundary of one flattened PV). Figures 6(b) and 6(c) show (on the same reference image) the tracker-recorded catheter positions of acquired spectra, colored by model-predicted lesion probability. Data in Fig. 6(b) were acquired in PBS, whereas data in Fig. 6(c) were acquired in blood. Predicted lesion probabilities match ground truth locations well, regardless of mapping environment. A probability threshold of 0.5 yielded the overall confusion matrix shown in Fig. 6(d), when applied to the entire test dataset. Applying this threshold individually to the PBS and blood test data yielded the confusion matrices in Figs. 6(e) and 6(f).

Fig. 6

Lesion classification results for an example heart in the test set. (a) Reference image showing the known locations of a line of 9 lesions generated near a pulmonary vein. (b) Predicted lesion probabilities from data collected with the heart submerged in PBS. (c) Predicted lesion probabilities from data collected with the heart submerged in blood. (d) Confusion matrix output for the trained RF model applied to the full test set, after applying a probability threshold of 0.5. (e) Confusion matrix output for (test set) PBS data. (f) Confusion matrix output for (test set) blood data. In panels (b) and (c), each point represents a single acquired spectrum.

3.3.

Blood Mapping Robustness

Data from all four test hearts were processed and individually compared as described in Sec. 2.5. The procedure for obtaining thresholded binary predictions and computing Dice coefficients is illustrated (for test heart 1) in Fig. 7. Calculated image comparison metrics are reported in Table 1. Across all test hearts, the mean SSIM value was 0.749, the mean inner product value was 0.867, and the mean RMS difference was 0.186. The mean Dice coefficients, for threshold values of 0.3, 0.5, and 0.7, were 0.571, 0.653, and 0.421. The Dice coefficient for test heart 3 with a threshold value of 0.7 is not listed because both the PBS- and blood-based predictions had no points with predicted probabilities above 0.7. The degree of similarity between maps obtained in PBS and blood, quantified by these metrics, suggests that NIRS measurements are equivalent in the two media.

Fig. 7

Example comparison between data collected from test heart 1, submersed in both PBS and blood. (a) From left to right: scattered model output (lesion probabilities); model output after interpolation onto a uniform grid; model output after thresholding with a probability value of 0.7. Scattered, grid-interpolated, and thresholded data are all visualized with a common color scale (a probability range of 0.5 to 1.0). (b) From left to right: thresholded predictions from PBS-mapped data compared with ground truth; thresholded predictions from blood-mapped data compared with ground truth; thresholded predictions from PBS- and blood-mapped data compared with one another.

Table 1

Results of direct comparisons between PBS- and blood-mapped test set hearts. The example heart shown throughout Figs. 6 and 7 is labeled as test heart 1. The dashes in the rightmost column for test heart 3 signify that at a probability threshold of 0.7, the model predicted no lesions for either the PBS or blood map.