Optical Spectroscopic Detection and Typing of Cardiac Amyloidosis

Cardiac amyloidosis (CA) ¹ is an increasingly recognized cause of heart failure, with two amyloid types (immunoglobin light chain, AL; and transthyretin, ATTR) accounting for ~95% of cases. The two types are usually mutually exclusive, confer different prognoses, and require different treatments, necessitating reliable typing. Histopathologic CA typing currently involves staining, laser-microdissection, and shotgun proteomics using liquid chromatography-tandem mass spectrometry (LC-MS/MS) ². Unfortunately, LC-MS/MS is technically complex with limited deployment, can have significant turn-around-times, and depletes the limited amount of tissue from endomyocardial biopsies (EMB). Herein, we present a new technology – combining infrared (IR) spectroscopic imaging and artificial intelligence (AI) ³ – to provide a non-destructive, rapid, and automated approach to CA detection and typing by optically detecting protein secondary structures⁴.

An overview of this approach and a comparison to the current state-of the art is presented in Figure A. We analyzed de-identified formalin-fixed paraffin-embedded (FFPE) human tissue samples (N_total=129, Mean Age= 69.6 (29–95), Male=80.3%, Female=19.7%, White=86.3%, Black=6.9%, Hispanic=3.1%, Asian=0.7%, Mid-Eastern=0.8%, Unknown/others=2.2% ) from Mayo Clinic (N=105) and Cedars Sinai (N=24), encompassing both EMB (N=79) and autopsy (N=50) tissue. The study was approved by the Institutional Review Boards at UIUC, Mayo Clinic and Cedars-Sinai. Whole-slide hyperspectral IR images (4 cm⁻¹ resolution, 750–4000 cm⁻¹, 25 μm pixels, transflection mode, using a Spotlight-400 microscope from Perkin Elmer, on MirrIR slides from Kevley Technology) and Congo Red-stained images (NanoZoomer Slide Scanner, Hamamatsu) were recorded. Amyloid deposits were identified by a characteristic shoulder in IR spectra ~1632 cm⁻¹ (red) and collagen (green) by absorption at 1240 cm⁻¹ (Figure B (i)). IR spectra were analyzed as second derivatives (SD) (Savitzky-Golay filter, convolution window=11 points) (Figure B (ii),) with intensities at 1632 cm⁻¹ and 1240 cm⁻¹ readily highlighting amyloid and collagen in myocardium, respectively (Figure C (i) and (ii)), without the need for staining or human recognition. Brightfield and cross-polarized microscopy images of CR-stained slides were evaluated by board-certified pathologists and annotated as the gold standard. SD data were used to train an artificial neural network (ANN1, with class labels and SD intensity at 1204, 1628, 1648, 1660 cm⁻¹ as inputs in MATLAB) that assigns a probability of each pixel belonging to one of three categories. A representative result (Figure C (iii)) shows clear correspondence between the classified and CR-stained images (Figure C (iv)). Quantitative segmentation accuracy is demonstrated by a receiver operating characteristic (ROC) curve, the area under the ROC curve (AUC=0.987, F1-score-0.922), and the confusion matrix (inset Figure D), demonstrating high overall accuracy (94.8%), sensitivity (90.3%), and specificity (93.2%). Figure E (i)–(iii), shows a representative EMB sample demonstrating the size independence of the method.

Figure. Rapid Diagnostic IR imaging Workflow for Amyloid Detection and typing.

A: Schematic overview of the current clinical workflow (top, black lines) and proposed workflow (bottom, green lines) that significantly reduces number of steps and correlates with clinical workflow on gold standard diagnoses (blue dashed line). B: Average mid-IR fingerprint spectra (1000–1800 cm⁻¹) of the three histologic classes, showing (i) spectra, offset for clarity, and (ii) Vector normalized second derivative spectra. C: Chemical, segmented, and clinical images from a representative autopsy sample, showing normalized second derivative intensity at (i) 1632 cm⁻¹ and (ii) 1240 cm⁻¹, (iii) Classified image (color coded for histologic classes), and (iv) Brightfield image of Congo Red-stained adjacent tissue section. D: ROC curves for (color-coded, red: amyloid, blue: cardiomyocyte, green:collagen) amyloid detection algorithm (validation set). The inset shows the confusion matrix, characterized by percent of target pixels. Quantitative metrics for the test include an AUC of 0.987, F1-score of 0.922, and overall accuracy of 94.8% for amyloids. E: Translation of the algorithm, showing a representative EMB sample with (i) second derivative intensity at 1632 cm⁻¹, (ii) classified image and (iii) Congo red-stained adjacent section. F: F1 score plot for all samples (training and validation). Maximum F1-score (0.986) is obtained for a threshold of 0.57. This operating point in ROC leads to an accuracy of 98.4%, specificity of 96.5% and sensitivity of 100% for AL type. G: ROC curve for 87 validation patients only. The inset shows the confusion matrix with metrics of the classification being an AUC of 0.974, an F1-score of 0.981, and overall accuracy of 97.7%. H: Correlation of H_patient (defined in main text) with Mean Beta Ratio (defined in main text), showing a clear separation of AL (red circles) and ATTR (blue circles) patients across the decision boundary. IR: Infrared, ROC: Receiver operator characteristic curve, AUC: Area under curve.

We developed second model (ANN2, with SD Intensity at 848, 900, 1048, 1204, 1340, 1392, 1548, 1624, 1628, 1632, 1648, 1660, 1688, 3076, 3284 cm⁻¹ as inputs) for pixels identified as amyloid with >90% probability by ANN1. ANN2 was trained on MS data from 42 patients and independently validated on 87 patients. Pixel level predictions were averaged as the AI determined probability of a patient having AL-type amyloidosis (H_patient) with an objective decision boundary (Figure F) for optimal algorithm performance (threshold=0.57). Thus, the recording of IR spectra and subsequent analyses by AI can be automated and objective, without the need for human intervention. At this operating point, high accuracy (97.7%), specificity (94.4%), sensitivity (100%), and positive predictive value (96.1%) were found for 87 patients in the independent validation. Notably, the model was trained primarily on EMB samples (N=39/42) from Mayo Clinic and validated on both EMB and autopsy samples from both clinical sources. Data and codes for training and validation, amyloid type data (predicted, ground truth, beta ratios) for all patients and patient demographics are available at https://figshare.com/projects/Optical_Spectroscopic_Detection_and_Typing_of_Cardiac_Amyloidosis/214765. Raw unprocessed data from 16 patient samples have been made available for demonstration purposes. Entire dataset will be made available on reasonable request. Two failure cases, wherein ATTR are misclassified as AL, had mitigating clinical conditions-multiple myeloma in the first and large background from polyclonal antibodies in the second. Finally, spectral deconvolution of the Amide I region was conducted to gain insight into the molecular basis of diagnoses by peak deconvolution for protein conformations - β-sheet, parallel or antiparallel (1630 cm⁻¹), random or β-turns (1660 cm⁻¹) or anti-parallel β−sheet (1690 cm⁻¹)-. A mean β-ratio (area of the fitted curves at 1630 cm⁻¹ to 1660 cm⁻¹) signified the relative proportion of β-sheet proteins to random coil or unresolvable structures, showing statistically significant higher value in ATTR (p=8.4e-12, two-sided unpaired t-test, Cohen’s D=1.5244) and segregating all cases (except the two failure cases in our model) (Figure H).

In summary we report a rapid, semi-automated, and non-destructive approach for cardiac amyloid detection and typing. Using only an IR microscope and AI, the technique can be broadly applied in clinical settings where a microscope can be housed, is straightforward to operate with minimal training and is compatible with current tissue processing and handling protocols. These features make the technology readily deployable into clinical laboratories that may have limited access to mass spectrometry or lack expertise for immunohistochemistry, while requiring significantly fewer consumables and labor. Finally, the non-perturbative imaging leaves the tissue intact for further analyses. These capabilities pave the way for its use as a clinical and research tool.

Nonstandard Abbreviations and Acronyms:

CA

Cardiac Amyloidosis

AL

Amyloid Light Chain

ATTR

Amyloid Transthyretin

LC-MS/MS

Liquid Chromatography Tandem Mass Spectrometry

IR

Infrared

EMB

Endomyocardial Biopsy

FFPE

Formalin Fixed Paraffin Embedded

ANN

Artificial Neural Network

SD

Second Derivative Spectra

ROC

Receiver Operator Characteristic

AUC

Area under curve