Systems and methods for diagnosis and prediction of respiratory infections using exhaled breath

Abstract

Systems and methods for diagnosis and prediction of respiratory infections include capturing and quantifying volatile organic compounds (VOCs) in exhaled breath samples. More specifically, detection of particular combinations of VOCs in atypical levels is indicative of the presence of a respiratory infection while detection of different combinations of VOCs in atypical levels is indicative of an elevated risk of future onset of a respiratory infection.

Description

[0001] SYSTEMS AND METHODS FOR DIAGNOSIS AND PREDICTION

[0002] OF RESPIRATORY INFECTIONS USING EXHALED BREATH CROSS-REFERENCE TO RELATED APPLICATION

[0003]

[0001] This application claims the benefit of priority to United States Provisional Patent Application Serial No. 63788004, filed April 13, 2025, for SYSTEMSAND METHODS FOR DIAGNOSIS AND PREDICTION OF RESPIRATORY INFECTIONS USING EXHALED BREATH, incorporated herein by reference.

[0004] FIELD OF THE INVENTION

[0005]

[0002] Systems and methods for diagnosis and prediction of respiratory infections include capturing and quantifying volatile organic compounds (VOCs) in exhaled breath samples. More specifically, detection of particular combinations of VOCs in atypical levels is indicative of the presence of a respiratory infection while detection of different combinations of VOCs in atypical levels is indicative of an elevated risk of future onset of a respiratory infection.

[0006] BACKGROUND OF THE INVENTION

[0007]

[0003] Postoperative (post-op) respiratory infections are a common complication following cardiac surgeries and have been associated with poor clinical outcomes and increased healthcare costs through longer stays and increased readmissions. Lower respiratory infections contributed to 3.2 million deaths globally in 2015. In 2021, pneumonia was the leading infectious cause of death, the 5th leading cause of death, and the leading cause of death in lower-income countries according to the World Health Organization. Furthermore, in younger populations, pneumonia has led to lung complications in adulthood including asthma, chronic obstructive pulmonary disease (COPD), and bronchiectasis. Early intervention of pneumonia has led to improved outcomes such as shorter hospital stay, shorter intensive care unit (ICU) stay, and lower rate of intensive care unit admission. Thus, anticipating pneumonia onset could improve patient outcome and significantly decrease clinical burden.

[0008]

[0004] Currently, chest radiography is one of the most common methods to diagnose pneumonia; however, notable challenges with this technique have indicated a need for improvement in diagnostic tools for pneumonia. Obstruction by anatomical structures, such as the heart, can make diagnosis difficult and has even led to falsenegatives being reported. Furthermore, analyzing images can be time-consuming, and interobserver agreement is usually poor to fair (K = 0.17-0.56). Recently, the analysis of volatile organic compounds (VOC) in diagnostic technology has provided a non-invasive method to diagnose patients through exhaled breath samples that is quick and less prone to interobserver disagreement. VOC analysis has already been used in investigating or diagnosing lung related conditions such as chronic obstructive pulmonary disease, asthma, tuberculosis, and cystic fibrosis. Furthermore, diagnostic tools have begun to utilize machine learning (ML) techniques for pattern recognition. With an increasing need of accuracy and timely diagnoses, usage of ML techniques could help alleviate these difficulties. Recently, VOCs have been evaluated through ML applications to discriminate between positive and negative SARS-CoV-2 samples, diagnose pneumonia, predict duration of prolonged mechanical intubation, and diagnose pulmonary fibrosis. Although several studies have used VOC profiles for varying diagnostic purposes, diagnosing pneumonia with VOCs remains challenging, with variance in the VOC profiles and metrics.

[0009]

[0005] A need exists for new systems and methods for diagnosis of respiratory infections by analysis of VOCs in exhaled breath samples with increased accuracy. An additional need exists for systems and methods to predict, not only detect, onset of respiratory infections.

[0010] SUMMARY

[0011]

[0006] The present invention addresses this need by providing systems and methods for diagnosis and prediction of respiratory infections by capturing and quantifying VOCs. More specifically, detection of particular combinations of VOCs in atypical levels is indicative of the presence of a respiratory infection, such as, for example, pneumonia, while detection of different combinations of VOCs in atypical levels is indicative of an elevated risk of contracting the respiratory infection. To this end, a machine learning workflow was implemented to analyze VOCs collected pre-operatively and at several timepoints after cardiac surgery. This invention offers the potential of early identification of patients at high risk of pneumonia following cardiac or other surgery, thus enabling predictions that could guide preventative care. In one embodiment, VOCs are collected from a subject patient and analyzed prior to thepatient undergoing a surgical procedure to help determine the patient’s post-op risk of respiratory infection.

[0012]

[0007] As compared to prior studies using VOC profiles for varying diagnostic purposes, the disclosed systems and methods utilize different breath capture technology, different methodology for evaluating VOCs, and are the first to demonstrate the ability to predict the future onset of pneumonia rather than only diagnosing existing cases of pneumonia. A medical services provider may modify the treatment of a patient based on this predicted risk. For example, if the disclosed systems and methods predict a specific patient to be at an elevated risk of developing post-operative pneumonia, that patient may be given more extensive prophylactic antibiotics in an effort to prevent the development of this respiratory infection. While it may not be not cost effective to give prophylactic antibiotics to all patients undergoing major operations, identifying a high risk group to intervene on, leading to the avoidance of this complication, would have significant health impacts for the patient, significant cost savings for the health care system, and efficiency in allocation of health care resources.

[0013]

[0008] In some embodiments, the present invention is a method for diagnosing a respiratory infection and predicting onset of the respiratory infection, the method comprising: capturing, using a capture device, an exhaled breath sample from the subject, wherein the exhaled breath sample includes a plurality of carbonyl-containing volatile organic compounds (VOCs); forming adducts of the plurality of carbonyl-containing VOCs with a reactive chemical compound; quantifying each of the adducts of each of the plurality of carbonyl-containing VOCs to establish a subject value for each of the adducts; comparing a first set of subject values to the dataset of standard values, wherein detection of the first set of subject values at decreased concentrations as compared to the standard values is indicative of elevated risk of onset of the respiratory infection; comparing a second set of subject values to the dataset of standard values, wherein detection of the second set of subject values at increased concentrations as compared to the standard values is indicative of elevated risk of onset of the respiratory infection; comparing a third set of subject values to the dataset of standard values, wherein detection of the third set of subject values at decreased concentrations as compared to the standard values is indicative of the presence of the respiratory infection; and comparing a fourth set of subject values to a dataset of standard values, wherein detection of the fourth set of subject values at elevatedconcentrations as compared to the standard values is indicative of the presence of the respiratory infection. In further embodiments, if the subject is identified as having elevated risk of onset of the respiratory infection, the method further includes administering prophylactic antibiotics to the subject. In certain embodiments, the respiratory infection is pneumonia. In some embodiments, the capturing occurs prior to a surgical procedure on the subject. In further embodiments, the first set of subject values comprises subject values of heptanal, octenone, and acetone. In certain embodiments, the second set of subject values comprises subject values of C12H24O. In some embodiments, the first set of subject values comprises subject values of heptanal, octenone, and acetone, and wherein the second set of subject values comprises subject values of C12H24O. In further embodiments, the third set of subject values comprises subject values of acrolein, and the fourth set of subject values comprises subject values of formaldehyde. In certain embodiments, the fourth set of subject values further comprises subject values of hexanone. In some embodiments, the third set of subject values comprises subject values of hexanal, and the fourth set of subject values comprises subject values of formaldehyde. In further embodiments, the fourth set of subject values further comprises subject values of C10H20O and C11H22O. In certain embodiments, the third set of subject values comprises subject values of at least one, at least two, at least three, at least four, at least five, at least six, or all of hexanal, acetaldehyde, acetone, butanal, acrolein, C7H11O, heptanal, hydroxy-2-butanone, crotonaldehyde, octenone, C5H8O, and C3H6O2, and wherein the fourth set of subject values comprises subject values of at least one, at least two, at least three, at least four, at least five, at least six, or all of formaldehyde, C10H20O, C11H22O, hexanone, hydroxy-butanal, 4-HHE, octanal, 4-HNE, 2-butanone, nonanal, C6H10O, C2H4O, pentanone, and C12H24O.

[0014]

[0009] In some embodiments, the present invention is a method for reducing the risk of post-operative respiratory infection, the method comprising: capturing, prior to a surgical procedure, an exhaled breath sample from the subject, wherein the exhaled breath sample includes a plurality of carbonyl-containing volatile organic compounds (VOCs); forming adducts of the plurality of carbonyl-containing VOCs with a reactive chemical compound; quantifying each of the adducts of each of the plurality of carbonyl-containing VOCs to establish a subject value for each of the adducts; comparing a first set of subject values to the dataset of standard values, wherein detection of the first set of subject values at decreased concentrations as comparedto the standard values is indicative of elevated risk of onset of the respiratory infection; comparing a second set of subject values to the dataset of standard values, wherein detection of the second set of subject values at increased concentrations as compared to the standard values is indicative of elevated risk of onset of the respiratory infection. In further embodiments, if the subject is identified as having elevated risk of onset of the respiratory infection, the method further comprises administering prophylactic antibiotics to the subject. In certain embodiments, if the subject is identified as having elevated risk of onset of the respiratory infection, the method further comprises identifying the subject as a candidate for receiving a prophylactic dose of antibiotics. In some embodiments, the respiratory infection in pneumonia. In further embodiments, the surgical procedure is a cardiac surgical procedure. In certain embodiments, the first set of subject values comprises subject values of heptanal, octenone, and acetone. In some embodiments, the second set of subject values comprises subject values of C12H24O. In further embodiments, the first set of subject values comprises subject values of heptanal, octenone, and acetone, and wherein the second set of subject values comprises subject values of C12H24O.

[0015]

[0010] It will be appreciated that the various systems and methods described in this summary section, as well as elsewhere in this application, can be expressed as a large number of different combinations and subcombinations. All such useful, novel, and inventive combinations and subcombinations are contemplated herein, it being recognized that the explicit expression of each of these combinations is unnecessary.

[0016] BRIEF DESCRIPTION OF THE DRAWINGS

[0017]

[0011] A better understanding of the present invention will be had upon reference to the following description in conjunction with the accompanying drawings.

[0018]

[0012] FIG. 1 depicts a schematic overview of the machine learning pipeline for diagnosis (nD = 75) and onset predictions (nR = 73) for pneumonia. Workflow, in order, of log transformation, normalization, and autoscaling is depicted in shaded boxes. New observations synthesized by oversampling techniques are depicted in shaded boxes marked with cross-hatching.

[0019]

[0013] FIG. 2A includes charts depicting the performance of the optimum model for diagnosing pneumonia. Specifically, FIG. 2A includes the area under the receiver operating characteristic curve (AUROC) (left) and area under the precision recall curve (PRAUC) (right) curves for the training (dashed line; AUROCTRAIN, PRAUCTRAIN) andtesting (solid line; AUROCTEST, PRAUCTEST) sets. AUROCTRAIN 95% confidence interval (Cl) ranged from 0.856-0.907. PRAUCTRAIN 95% Cl ranged from 0.847-0.910.

[0020]

[0014] FIG. 2B includes charts depicting the performance of the optimum model for predicting the onset of pneumonia. Specifically, FIG. 2B includes AUROC (left) and PRAUC (right) curves for the training (dashed line) and testing (solid line) sets. AUROCTRAIN 95% Cl ranged from 0.872-0.939. PRAUCTRAIN 95% Cl ranged from 0.874-0.956.

[0021]

[0015] FIG. 3A is a chart depicting partial least squares discrimination analysis (PLSDA) of the training and testing sets for diagnosing pneumonia.

[0022]

[0016] FIG. 3B is a chart depicting PLSDA of the training and testing sets for predicting onset of pneumonia.

[0023]

[0017] FIG. 4 is a series of bar graphs depicting the maximum training AUROC for each model across all feature subsets for each sampling method for diagnosing pneumonia (left) and predicting onset (right). Each row of graphs showcases the performance of each model during training using the different oversampling methods. The optimum models were determined to be random forest (RF) using synthetic minority oversampling technique (SMOTE) for diagnosing pneumonia and boosted general linear model (GLMB) using k-means clustering SMOTE for predicting onset of pneumonia, as indicated by darker shading in the figure. Both models obtained the highest AUROCTRAIN relative to the other models in their respective runs after filtering for AUROCTRAIN >0.70, AUROCTEST >0.70, PRAUCTRAIN >0.70, PRAUCTEST >0.70, FITRAIN > 0.70, FITEST >0.70 and AUROCTEST < AUROCTRAIN to prevent overfitting, thus being selected over the remaining models (light grey) with weaker performances. AVNNET = neural networks using model averaging. PCANNET = neural networks with principal component step. GLM = generalized linear model. GLMB = boosted generalized linear model. KNN = K-nearest neighbors. NB = naive Bayes. NNET = neural network (single layer). PAM = nearest shrunken centroids. SVMRW = support vector machine with radial basis function kernel and class weights. SVML = support vector machine using linear decision boundary. SVMR = support vector machine radial.

[0024]

[0018] FIG. 5A includes a chart depicting performance scores for each feature subset of RF for pneumonia diagnostics. The dashed vertical line denotes the number of features used by the optimum model (i.e., 21 features).

[0019] FIG. 5B includes a chart depicting performance scores for each feature subset of GLMB for predicting onset. The dashed vertical line denotes the number of features used by the optimum model (i.e., 4 features).

[0025] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0026]

[0020] The details of one or more embodiments of the presently-disclosed subject matter are set forth in this document. Modifications to embodiments described in this document, and other embodiments, will be evident to those of ordinary skill in the art after a study of the information provided in this document. The information provided in this document, and particularly the specific details of the described exemplary embodiments, is provided primarily for clearness of understanding and no unnecessary limitations are to be understood therefrom. In case of conflict, the specification of this document, including definitions, will control.

[0027]

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the presently-disclosed subject matter belongs. Although any methods, devices, and materials similar or equivalent to those described herein can be used in the practice or testing of the presently-disclosed subject matter, representative methods, devices, and materials are now described.

[0028]

[0022] Following long-standing patent law convention, the terms “a”, “an”, and “the” refer to “one or more” when used in this application, including the claims. Unless otherwise indicated, all numbers expressing quantities of ingredients, properties such as reaction conditions, and so forth used in the specification and claims are to be understood as being modified in all instances by the term “about.” Accordingly, unless indicated to the contrary, the numerical parameters set forth in this specification and claims are approximations that can vary depending upon the desired properties sought to be obtained by the presently-disclosed subject matter. As used herein, the term “about,” when referring to a value or to an amount is meant to encompass variations of ±10% of the most precise digit in the value or amount (e.g., “about 1” refers to 0.9 to 1.1, “about 1.1” refers to 1.09 to 1.11, etc.).

[0029]

[0023] As used herein, ranges can be expressed as from “about” one particular value, and / or to “about” another particular value. It is also understood that there are a number of values disclosed herein, and that each value is also herein disclosed as “about” that particular value in addition to the value itself. For example, if the value “10” isdisclosed, then “about 10” is also disclosed. It is also understood that each unit between two particular units are also disclosed. For example, if 10 and 15 are disclosed, then 11, 12, 13, and 14 are also disclosed.

[0030]

[0024] Methods

[0031]

[0025] De-identified cardiac surgery patient data were obtained from patients undergoing elective cardiac surgery at the University of Louisville. The dataset included samples of patient volatile organic compounds (VOC) profiles from before surgery (baseline), directly after surgery, and samples taken in 3-day intervals postsurgery until discharge. Pneumonia diagnosis was evaluated by a qualified physician at each interval.

[0032]

[0026] To collect the sample for VOC analysis, patients exhaled through a filter (Niox-Vera, Morrisville, NC) connected to a Teflon tube via a 3D-printed coupling. Two separate Tedlar bags (Sigma-Aldrich, St Louis, MO) were attached to the apparatus via a valve. The valve was turned after filling the first bag (150 milliliters), allowing breath into the second bag (0.6 L). The first bag was discarded (as dead space) and the second bag was used for VOC analysis.

[0033]

[0027] Breath samples were evacuated through a silicon microreactor consisting of an array of thousands of micropillars. Micropillar surfaces were coated with a quaternary ammonium compound, 2-(amino-oxy)-N, N, N trimethylethanammonium (ATM) triflate, synthesized from ATM iodide using a counterion exchange protocol. The ATM selectively reacts through oximation reactions with the sample and traps carbonyl compounds with capture efficiencies of 90% or greater at a flow rate of 7 mL / min. Procedures for using known microreactors to capture carbonyl VOCs in exhaled breath, identification of chemical structures of detected carbonyl compound, and the quantitative analysis of ATM-carbonyl adducts by UHPLC-MS are known in the art. ATM-carbonyl adducts in the microreactor were eluted with 200 pL LC / MS grade water from a slightly pressurized vial to an empty collecting sample vial, with 99% recovery of ATM adducts. A known amount of deuterated acetone that had completely reacted with ATM (ATM-acetone-d6 adduct) was added to the eluted solution as internal reference. An ultra-high performance liquid chromatography-mass spectrometer (UHPLC-MS, Thermo Scientific) equipped with an autosampler, a Vanquish UHPLC, and a Q Exactive Focus Orbitrap Mass Spectrometer (MS) was used for analysis. VOC concentrations were determined by comparing relative abundance with that of added internal reference ATM-acetone-d6. Chemical structures of carbonyl compounds weredetermined by comparison of retention time and MS / MS spectra with the synthesized standard adducts.

[0034]

[0028] To identify pneumonia, patients were grouped by whether pneumonia developed (P+) or did not develop (P-) postoperatively over the course of hospital stay. VOC samples used for P+ patient samples were taken from the first sample associated with a positive pneumonia diagnosis, while P- patient samples were taken from the day before the mean P+ onset day for pneumonia.

[0035]

[0029] To pre-operatively predict onset of pneumonia, a second dataset was used comprising each patient’s baseline (pre-operative) sample. Patients that developed pneumonia during their hospital stay were classified as “High” risk and those that did not develop pneumonia were “Low” risk. Two patients were missing baseline samples and were excluded from predictions of disease onset.

[0036]

[0030] Uninduced biological variation can often introduce heteroscedasticity, where variance along a regression is not constant. To reduce the effects of heteroscedasticity of VOCs, a log transformation was applied to each dataset. Moreover, mass spectrometry total useful signal (TUS) normalization was performed to control for variation in sample concentrations.

[0037]

[0031] In both sets of data, the number of high-risk patients and P+ samples was smaller compared to low-risk and P- patients, respectively. To compensate for this imbalance, both datasets underwent a similar data-wrangling pipeline involving sampling techniques. The data for the low-risk patients and P- (majority class) were undersampled using random undersampling (RUS). RUS is a technique that removes samples from the majority class to reduce imbalance in a dataset, which has seen success in other clinical studies. The data for the high-risk patients and P+ (minority class) were randomly partitioned into a training set and a test set at a 7:3 ratio. Data sets for both diagnosis and onset classification included patients in the majority class at a 2:1 ratio to the minority class. The mean and standard deviation of the training set were used to autoscale both data sets to control for variance between features and emphasize differences within the data. To assess the heterogeneity between the training and test sets, prior to oversampling, partial least squares discrimination analysis (PLSDA) was performed using mdatools (v0.14.1) package.

[0038]

[0032] To provide adequate number of training samples for the high-risk patients and P+ patients, the training set for each classification was oversampled using synthetic minority oversampling technique (SMOTE) or variations of SMOTE. SMOTE and itsvariants are oversampling techniques that generate new samples of the minority class to reduce imbalance in the dataset. These techniques have been utilized in other clinical datasets with success, showing increased performance in ML training. Using the themis (v1.0.2) package, base SMOTE was used for diagnosis classification while k-means clustering SMOTE was used for onset prediction. To ensure oversampling did not bias model generalizability, SMOTE techniques were not applied to the test set.

[0039]

[0033] To predict diagnosis and onset of pneumonia, each dataset was input to a ML workflow that included forward-feature selection and model training. Training and testing were conducted using caret (v6.0-94) package. Hyperparameter tuning was performed using a grid search with Cohen’s kappa as the optimization metric with a tune length of 10. Model performance estimates were approached using a 5-fold cross validation and 20 resampling iterations for each base learner. Area under the receiver operating characteristic curve (AUROC) scores were measured from 0-1, where a score of 1 indicates perfect predictive performance and a score of 0.5 represents the equivalent of random chance. F1 scores — a metric that balances precision and recall — were used in combination with AUROC to quantify model performance. Each model’s top feature ranking was computed using the varlmp function, provided by caret, to show variable importance. AUROC score and confidence interval (Cl) were computed and visualized with pROC (v1.18.5) using the DeLong method. Area under the precision recall curve (PRAUC) was computed and visualized with the pr.curve function in PRROC (v1.3.1) and the Cl was computed by following known methods then applying 10,000 replicates of bootstrapping using boot (v1.3-31) and finding the mean values of PRAUC and maximum F1 scores.

[0040]

[0034] Machine learning models evaluated in the study included principal component step (pcaNNet), model averaged neural network (avNNet), multi-layer perceptron (mlpWeightDecayML), k-nearest neighbor (knn), random forest (rf), neural networks (nnet), generalized linear model (glm), naive bayes (naive_bayes), boosted generalized linear model (glmboost), pam, support vector machine with class weights (svmRadialWeights), support vector machine with linear kernel (svmLinear), and support vector machine with radial basis function (svmRadial).

[0041]

[0035] A relative abundance analysis was performed on each dataset to compare sample groups using T-test or Wilcoxon rank sum test on parametric and nonparametric data, respectively, as determined by Shapiro-Wilks test. To understand thecorrelation between VOCs and the onset of pneumonia, a Spearman’s test was conducted for each VOC against each outcome to find any significant correlation coefficients, where strong correlation was defined as |r| > 0.75. Analysis was performed in R (v4.3.1).

[0042]

[0036] Results

[0043]

[0037] The details of the patient population are summarized in Table 1. As seen in FIG.

[0044] 1, the schematic overview of the data wrangling pipeline for diagnosis (nD = 75) and onset predictions selected 73 patients for analysis. As used herein, “data wrangling” refers to transforming raw data into a different format more useful for downstream purposes, such as analytics. Among the diagnostic samples, only formaldehyde and acrolein (p < 0.05) were found to significantly differ between P+ and P- samples (Table 2). Among the baseline samples, only heptanal (p < 0.05) was found to significantly differ between low- and high-risk patients (Table 3). For the diagnostic samples, the mean number of days until onset was 5.4 days for P+ patients.

[0045] Table 1: Patient Population Characteristics

[0046] .... Pneumonia Developed,, Characteristic - P-value? Age _ 63 (52, 71) 66 (64, 69) _ 0.197 _

[0047] Gender >0.900

[0048] F _ 18 (28%) 3 (30%) _

[0049] M _ 47 (72%) 7 (70%) _

[0050] 1Median (IQR); n (%).2Wilcoxon rank sum test; Fisher’s exact test. P- = did not develop pneumonia during hospitalization. P+ = developed pneumonia during hospitalization.

[0051] Table 2: Patient Sample Characteristics for Diagnosing Pneumonia Pneumonia Diagnosis

[0052] Characteristic - - - - - p-value2

[0053] P-, N = 657P+, N = 107

[0054] 0.0408 0.0417 Formaldehyde _ _ 0.033

[0055] J(0.0403, 0.0415) (0.0412, 0.0424)

[0056] . 0.0388 0.0384 „ „„ Acetaldehyde 0.137y(0.0384, 0.0391) (0.0382, 0.0387)

[0057] 0.0426 0.0418 Acetone 0.362

[0058] (0.0411, 0.0439) (0.0410, 0.0432)

[0059] „ 0.0359 0.0359

[0060] 2-Butanone >0.900

[0061] 354, 0.0361 )

[0062] 61335

[0063] U ana

[0064]

[0065] 334, 0.0338)00338 0.0340

[0066] PentanOne(0.0333, 0.0341) (0.0340, 0.0343) °1 17„r, 0.0335 0.0333

[0067]

[0068] en ana(0.0331, 0.0339) (0.0331, 0.0338) 0.0314 0.0316 „ „ „HeXan°ne(0.0308, 0.0317) (0.0313, 0.0319) °17300344 00343

[0069]

[0070] (0.0340, 0.0347) (0.0342, 0.0345) 0.0303 0.0305

[0071] Heptanone 0.809 (0.0297, 0.0309) (0.0297, 0.0307)

[0072] 6.0332 777ep ana(0.0330, 0.0337) (0.0329, 0.0335)

[0073] 00347 0.0346

[0074] Octanal _ _n_ _ _ 0.895 (0.0344, 0.0350) (0.0345, 0.0350)

[0075] 0379

[0076] Nonana _ _ _ _ 0.133 (0.0373, 0.0382) (0.0377, 0.0386)

[0077] 0.0362 0.0363

[0078] C10H20O 0.528 (0.0357, 0.0368) (0.0358, 0.0369)

[0079]

[0080] (0.0313, 0.0336) (0.0303, 0.0334)

[0081] 1Median (IQR); n (%. ^Wilcoxon rank sum test; Fisher’s exact test. P- = did not develop pneumonia during hospitalization. P+ = developed pneumonia during hospitalization.

[0082] Table 3: Baseline Sample Characteristics for Stratifying Risk Pneumonia Risk

[0083] Characteristic p-value2

[0084]

[0085] -0 28 0 02

[0086]

[0087] i_i, i 0.00 -0.87 _nno-.

[0088] Heptana(-0.52, 0.47) (-1.16, -0.44) 0.0087

[0089] _Qctanal_ (-0.61, 0.65) (-0.93, 0.10) _0, 103

[0090] Nonana(-0.52, 0.44) (-0.61, 0.17)0,356

[0091] _CioH2oQ_ (-0,53,0.64) (-0,32, 0.65) _0,791

[0092] r u n 0.19 0.15

[0093] _Ci i H22° _ (-0.48, 0.64) (-0.32, 0.50) _0,79

[0094] r u n 0.09 0.00

[0095] Cl2H24° (-0.51, 0.73) (-0.09, 0.36)0,6 3

[0096] _C2H4°2_ (-0.80, 0.68) (-0.48, 0.88) _0,475

[0097] C3H6°2(-0.26, 0.72) (-1.35, 0.37)0, 130

[0098] Hydroxy-2-butanone (.p.^s t) (-0.65.^. SO)0,779

[0099] ..., -0.19 0.46

[0100] Hydroxy-butanaln

[0101] 0 5c

[0102] 8on

[0103] 0

[0104]

[0105] _Acrolein_ (-0.66, 0.61 ) (-0.30, 0.52) _0,475

[0106] Crotonaldehyde (-0.60, 0.44) (-0.99, 0.52)0,485

[0107] -0.06 -0.28

[0108] C5H8O 0.841

[0109] -0.79, 0.50) (-1.13, 1.18r u n -0.05 ■°'10n AQZI _CsHloQ_ (-0.83, 0.80) (-1.17, 0.69) _0 694

[0110] Octenone (-0.71, 0.87) (-0.96, 0.33)0 232

[0111] _C9Hl6° _ (-0.73, 0.75) (-0.70, 0.93) _0 251

[0112] Mn« -0.120 36n ncv

[0113] MDA(-0.62, 0.80) (-0.03, 1.11) 0.0573

[0114] 4 HHF0 14 0 280279

[0115] _4~HHE_ (-0.50, 0.52) (0.10, 0.51) _0 279

[0116] . uMr -0.12 -0.46n on.

[0117] 4-HNE(-0.67, 0.39) (-0.83, 0.97) °'804

[0118] 0.09 0.08

[0119] CzHnO 0.791

[0120] -0.32, 0.54) (-0.41, 0.79)

[0121] 1Median (IQR); n (%j.2Wilcoxon rank sum test; Fisher’s exact test.

[0122] Risk categorization was dependent on the patient developing pneumonia during hospitalization. “Low” = did not develop pneumonia during hospitalization. “High” = developed pneumonia during hospitalization.

[0123]

[0038] Correlation coefficients for patients’ exhaled VOC profiles were computed to identify the relationship between VOC concentration and presence of pneumonia and VOC concentration and onset of pneumonia. Using the samples for diagnosing pneumonia, no VOC was found to have moderate or strong correlations (Table 4). Using the baseline samples for stratification, only heptanal was found to have a moderate ( |r| > 0.30) negative correlation with onset of pneumonia (Table 5). No strong correlations (|r| > 0.75) were found among the VOCs in relation to either outcome, suggesting that statistical analyses alone would be insufficient to associate VOCs with pneumonia diagnosis or onset.

[0124] Table 4: Relative abundance of VOCs for diagnosing pneumonia voc. Direction of

[0125] p

[0126] r-value

[0127] Change

[0128] Formaldehyde ** 0.033 Increase

[0129] Acrolein ** 0.049 Decrease

[0130] Hexanone * 0.090 Increase

[0131] Pentanone 0.104 Increase

[0132] C11H22O 0.105 Increase

[0133] Nonanal 0.133 Increase

[0134] Acetaldehyde 0.137 Decrease

[0135] C12H24O 0.137 Increase

[0136] 4-HNE 0.173 Increase

[0137] Octenone 0.187 Decrease

[0138] MDA 0.210 Increase

[0139] Hydroxy-2-Butanone 0.221 Decrease

[0140] Hydroxy-Butanal 0.227 Increase

[0141] Butanal 0.265 Decrease

[0142] C2H4O2 6.359 increase_ 4-HHE _ 0.435 Increase

[0143] CgHieO 0.450 Increase

[0144] _ C10H20O _ 0.542 Increase

[0145] Pentanal 0.602 Decrease

[0146] _ Heptanal _ 0.608 Decrease

[0147] Acetone 0.609 Decrease

[0148] _ C3H6O2_ 0.668 Decrease

[0149] C7H11O 0.703 Increase

[0150] _ CsHioO _ 0.707 Increase

[0151] 2-Butanone 0.743 Decrease

[0152] _ Hexanal _ 0.755 Decrease

[0153] Octanal 0.821 Decrease

[0154] Crotonaldehyde _ 0.863 Increase

[0155] Heptanone 0.899 Increase

[0156] C5H8O 0.977 Increase

[0157] Direction of change is relative to the change in VOC median area when moving from the “N” (P-) to the “Y” (P+) groups. “Increase” indicates VOC median area is greater in the P+ groups as compared to the P- groups, while “Decrease” indicates VOC median area is lower in the P+ groups as compared to the P- groups. ** = significant (p <0.05). * = marginally significant (p <0.1 ).

[0158] Table 5: Relative abundance of VOCs in terms of pneumonia onset

[0159] 1 Direction

[0160] _ VOC P~Va Ueof Change

[0161] Heptanal * _ 0.009 Decrease

[0162] MDA * 0.026 Increase

[0163] Acetone 0.057 Increase

[0164] C3H6O2_ 0.129 Decrease

[0165] Octanal 0.144 Degrease

[0166] Pentanal _ 0.202 Decrease

[0167] _ Octenone 0.229 Decrease

[0168] Hexanal 0.238 Decrease

[0169] _ CgHisO _ 0.251 Increase

[0170] Heptanone 0.254 Decrease

[0171] _ 4-HHE _ 0.279 Increase

[0172] Nonanal 0.356 Decrease

[0173] _ C2H4O2 _ 0.368 Increase

[0174] Acetaldehyde 0.456 Increase

[0175] _ Hexanone _ 0.474 Decrease

[0176] Acrolein 0.475 Increase

[0177] Crotonaldehyde _ 0.485 Decrease

[0178] Formaldehyde 0.547 Increase

[0179] Pentanone _ 0.561 Increase

[0180] Hydroxy-Butanal 0.58 Increase

[0181] _ Butanal _ 0.602 Increase

[0182] C12H24O 0.613 Decrease

[0183] _ CsHioO _ 0.756 Decrease

[0184] 4-HNE 0.772 Decrease

[0185] Hydroxy-2-Butanone 0.779 Increase

[0186] C10H20O 0.791 IncreaseC11 H22O _ 0.791 Decrease

[0187] C7H11O 0.791 Decrease

[0188] 2-Butanone 0.885 Increase

[0189] C5H8O 0.891 Decrease

[0190] Direction of change is relative to the change in VOC median area when moving from the low-risk to the high-risk groups. “Increase” indicates VOC median area is greater in the high-risk groups as compared to the low-risk groups, while “Decrease” indicates VOC median area is lower in the high-risk groups as compared to the high- risk groups. * = significant (p <0.05).

[0191]

[0039] Before oversampling, PLSDA was used to compare the training and test sets for both outcomes. Under both outcomes, the training and test sets had significant differences, confirming heterogeneity between the sets (p < 0.05; Table 6).

[0192] Table 6: Population characteristics (n = 33) comparison for stratification between training and test set before oversampling.

[0193] _ Diagnosis of Pneumonia

[0194]

[0195] Predictign i of Pneumonia Onset

[0196]

[0197] _ Set _

[0198]

[0199] _ Set _

[0200] Characteristic Train, N = Test, N = 97p-value2Train, N = Test, N = P- 4 24797value 27 2

[0201]

[0202]

[0203]

[0204] 0.20 0.61 0.24 -0.01 CzHnO (-0.85, 0.74) (0.30, 0.90) 0.238 (-0.27, (-0.62, 0.706

[0205]

[0206]

[0207]

[0208] 0.62) 1.67)

[0209] 1n (%) or median (IQR).2Wilcoxon rank sum test.

[0210]

[0040] To diagnose pneumonia in patients following elective cardiac surgery, pneumonia positive (P+) VOC samples from the day of disease onset and pneumonia (P-) VOC samples closest to the mean P+ onset day were passed into a ML workflow. Models were trained to classify patients into P+ and P- groups. Considering both training and testing results, model training reported that the optimum validation subset AUROC performance was achieved by random forest using 21 features after application of SMOTE (FIG. 4). Internal validation reported an AUROCTRAIN 0.85, PRAUCTRAIN > 0.85, and F1 TRAIN > 0.75. The model was then applied to the test set, achieving an AUROCTEST 0.80, PRAUCTEST 0.80, and F1TEST > 0.80.

[0211]

[0041] To assess the potential of predicting disease onset, baseline VOC samples obtained before surgery were passed into the same ML workflow. Models were trained to make classifications on patients in “Low” risk (did not develop pneumonia) or “High” risk (developed pneumonia) groups. Model training reported that the optimum validation subset AUROC performance was achieved by a boosted generalized linear model using 4 features after application of k-means clustering SMOTE (FIG. 4). Internal validation reported an AUROCTRAIN > 0.90, PRAUCTRAIN > 0.90, and F1 TRAIN > 0.85. The model was then applied to the test set, achieving an AUROCTEST ^0.80, PRAUCTEST >0.80, and F1 TEST

[0212]

[0213] 0.80. The training and test performance curves for diagnosis and onset are shown in FIGs. 2A and 2B, respectively. Models for diagnostic and onset predictions yielded differing outcome probabilities for test set predictions that mapped to identical ROC and PR curves (note that the curves in FIGs. 2A and 2B denote ROC and PR whereas the areas under the curves are AUROC and PRAUC, respectively). Performance metrics across all feature subsets for both outcomes are visualized in FIGs. 5A and 5B.

[0214]

[0042] Features from the optimum models were ranked in order of importance (Table 7). 24 features were used to screen for pneumonia diagnosis including several aldehydes such as nonanal, heptanal, and octanal; ketones such as hexanone, acetone, or butanone; and other hydrocarbons. Four features were used to classify disease onset, primarily using select ketones and aldehydes.

[0215] Table 7: Feature Selection (FS) as Ranked by the Optimum Model via ML Workflow

[0216]

[0217] Higher FS scores imply increased importance to model predictions. Positive correlation indicates increased concentration of a given VOC was associated with presence of pneumonia. * = included in the optimum model.

[0218]

[0043] Altogether, these results indicate that VOC breath data from elective cardiac surgery patients is indicative of pneumonia diagnosis and is predictive of pneumoniaonset. In some embodiments, decreased concentration of a first set of VOCs, namely, heptanal, octenone and acetone and increased concentration of a second set of VOCs, namely, C12H24O, are biomarkers of future onset of pneumonia (taken from Table 7). In some embodiments, decreased concentrations of a third set of VOCs, namely, hexanal, acetaldehyde, acetone, butanal, acrolein, C7H11O, heptanal, hydroxy-2-butanone, crotonaldehyde, octenone, C5H8O, and C3H6O2and increased concentrations of a fourth set of VOCs, namely, formaldehyde, C10H20O, C11H22O, hexanone, hydroxy-butanal, 4-HHE, octanal, 4-HNE, 2-butanone, nonanal, C6H10O, C2H4O, pentanone and C12H24O are biomarkers indicating the presence of pneumonia (taken from Table 7). In other embodiments, the third set includes or consists of acrolein and the fourth set includes or consists of formaldehyde and hexanone (focusing on the VOCs of marginal significance or better in Table 4; p < 0.1 ). In further embodiments, the third set includes acrolein and the fourth set includes formaldehyde (focusing on the VOCs of significance in Table 4; p < 0.05). In other embodiments, the third set includes or consists of hexanal and the fourth set includes or consists of formaldehyde (focusing on the VOCs with the top 2 FS scores in Table 7. In further embodiments, the third set includes or consists of hexanal and the fourth set includes or consists of formaldehyde, C10H20O and C11H22O (focusing on the VOCs with the top 4 FS scores in Table 7).

[0219]

[0044] This disclosure shows consistent performance across train and test sets that were distinct from one another under PLSDA analysis (FIGs. 3A and 3B), which showcases model generalizability. This analysis indicates that pneumonia in patients undergoing elective cardiac surgeries is associated with unique VOC characteristics present both prior to surgery and at diagnosis.

[0220]

[0045] While the examples herein discuss using panels of VOCs captured and quantified pre-op as biomarkers for the diagnosis or onset of pneumonia, the same techniques may be used to identify panels of VOCs as biomarkers for the diagnosis or onset of other respiratory diseases. Also, while the specification discusses collection of VOC samples from exhaled breath using a specific capture device, other breath collection devices such as, for example, the breath collection tool disclosed in PCT / US2023 / 085017, may also be used.

[0221]

[0046] The foregoing detailed description is given primarily for clearness of understanding and no unnecessary limitations are to be understood therefrom, formodifications can be made by those skilled in the art upon reading this disclosure and may be made without departing from the spirit of the invention.

Claims

CLAIMS1. A method for diagnosing a respiratory infection and predicting onset of the respiratory infection, the method comprising:capturing, using a capture device, an exhaled breath sample from the subject, wherein the exhaled breath sample includes a plurality of carbonyl-containing volatile organic compounds (VOCs);forming adducts of the plurality of carbonyl-containing VOCs with a reactive chemical compound;quantifying each of the adducts of each of the plurality of carbonyl-containing VOCs to establish a subject value for each of the adducts;comparing a first set of subject values to the dataset of standard values, wherein detection of the first set of subject values at decreased concentrations as compared to the standard values is indicative of elevated risk of onset of the respiratory infection; comparing a second set of subject values to the dataset of standard values, wherein detection of the second set of subject values at increased concentrations as compared to the standard values is indicative of elevated risk of onset of the respiratory infection;comparing a third set of subject values to the dataset of standard values, wherein detection of the third set of subject values at decreased concentrations as compared to the standard values is indicative of the presence of the respiratory infection; and comparing a fourth set of subject values to a dataset of standard values, wherein detection of the fourth set of subject values at elevated concentrations as compared to the standard values is indicative of the presence of the respiratory infection.

2. The method of claim 1, further comprising, if the subject is identified as having elevated risk of onset of the respiratory infection, administering prophylactic antibiotics to the subject.

3. The method of claim 1, wherein the respiratory infection is pneumonia.

4. The method of claim 1, wherein the capturing occurs prior to a surgical procedure on the subject.

5. The method of claim 1, wherein the first set of subject values comprises subject values of heptanal, octenone, and acetone.

6. The method of claim 1, wherein the second set of subject values comprises subject values of C12H24O.

7. The method of claim 1, wherein the first set of subject values comprises subject values of heptanal, octenone, and acetone, and wherein the second set of subject values comprises subject values of C12H24O.

8. The method of claim 1, wherein the third set of subject values comprises subject values of acrolein, and wherein the fourth set of subject values comprises subject values of formaldehyde.

9. The method of claim 8, wherein the fourth set of subject values further comprises subject values of hexanone.

10. The method of claim 1, wherein the third set of subject values comprises subject values of hexanal, and wherein the fourth set of subject values comprises subject values of formaldehyde.

11. The method of claim 10, wherein the fourth set of subject values further comprises subject values of C10H20O and C11H22O.

12. The method of claim 1, wherein the third set of subject values comprises subject values of hexanal, acetaldehyde, acetone, butanal, acrolein, C7H11O, heptanal, hydroxy-2-butanone, crotonaldehyde, octenone, C5H8O, and C3H6O2, and wherein the fourth setof subject values comprises subject values of formaldehyde, C10H20O, C11H22O, hexanone, hydroxy-butanal, 4-HHE, octanal, 4-HNE, 2-butanone, nonanal, C6H10O, C2H4O, pentanone, and C12H24O13. A method for reducing the risk of post-operative respiratory infection, the method comprising:capturing, prior to a surgical procedure, an exhaled breath sample from the subject, wherein the exhaled breath sample includes a plurality of carbonyl-containing volatile organic compounds (VOCs);forming adducts of the plurality of carbonyl-containing VOCs with a reactive chemical compound;quantifying each of the adducts of each of the plurality of carbonyl-containing VOCs to establish a subject value for each of the adducts;comparing a first set of subject values to the dataset of standard values, wherein detection of the first set of subject values at decreased concentrations as compared to the standard values is indicative of elevated risk of onset of the respiratory infection; comparing a second set of subject values to the dataset of standard values, wherein detection of the second set of subject values at increased concentrations as compared to the standard values is indicative of elevated risk of onset of the respiratory infection; andif the subject is identified as having elevated risk of onset of the respiratory infection, administering prophylactic antibiotics to the subject.

14. The method of claim 13, wherein the respiratory infection in pneumonia.

15. The method of claim 13, wherein the surgical procedure is a cardiac surgical procedure.

16. The method of claim 13, wherein the first set of subject values comprises subject values of heptanal, octenone, and acetone.

17. The method of claim 13, wherein the second set of subject values comprises subject values of C12H24O.

18. The method of claim 13, wherein the first set of subject values comprises subject values of heptanal, octenone, and acetone, and wherein the second set of subject values comprises subject values of C12H24O.

Images