System and method for characterizing an aroma in a high-interference environment by disentangling confounders

Abstract

Systems and methods for disentangling confounders from volatile organic compound data with adversarial learning are described, for characterizing aroma in a high interference environment. In some implementations, sensor data including volatile organic compound signals and confounder-related signals may be received and encoded into a task-relevant latent representation and a confounder-related latent representation. The encoding may be performed by an encoder configured to separate the volatile organic compound signals from the confounder-related signals. A confounder predictor may be adversarially trained to extract the confounder-related signals from the confounder-related latent representation while the encoder may be trained to suppress the confounder-related signals in the task-relevant latent representation. The sensor data may be reconstructed from the task-relevant latent representation and the confounder-related latent representation using a decoder configured to preserve the volatile organic compound signals and isolate the confounder-related signals.

Description

[0001] STR-027W001

[0002] SYSTEM AND METHOD FOR CHARACTERIZING AN AROMA IN A HIGH- INTERFERENCE ENVIRONMENT BY DISENTANGLING CONFOUNDERS

[0003] FIELD OF TECHNOLOGY

[0004] [1] The present disclosure relates to a platform for characterizing an aroma, including identifying and quantifying the constituting gases or Volatile Organic Compounds of the aroma, in an environment that contains interfering analytes or parameters.

[0005] BACKGROUND

[0006] [2] Digital olfaction systems, or electronic noses (e-noses), have emerged as a promising technology for non-invasive diagnostics by mimicking the human sense of smell to detect patterns of aroma, which is defined as a mixture of volatile organic compounds (VOCs) and inorganic compounds. E-noses are based on an sensor array of partially selective sensing elements with cross-reactivity and use pattern-recognition methods to interpret complex aroma mixtures. Several e-nose architectures have been developed, differentiated primarily by their transduction mechanisms. The mam types currently used include chemiresistive, piezoelectric, optical, electrochemical, bioelectronic or hybrid systems.

[0007] Chemiresistive e-noses are the most widely employed due to their simplicity, scalability, and compatibility with portable designs. These systems rely on materials whose electrical resistance changes upon VOC or inorganic gas interaction, such as absorption, examples of these materials are conductive polymers, functionalized polymer based composite, polymercarbon black composites, metal-oxide semiconductors, carbon nanostructures such as carbon nanotubes (CNTs), graphene or any other allotrope of carbon.

[0008] [3] E-nose technology presents the potential to analyze exhaled breath for early detection of diseases such as cancer and metabolic disorders, offering a rapid, cost-effective, and patient-friendly alternative to invasive methods. These systems operate by capturing a characteristic “scentprint” from a sample’s VOC profile that can reflect underlying physiology. Scentprint is defined as the temporal response / time series data of the plurality of partially selective sensing elements of the enose sensor array generated when the array interacts with an incoming or introduced aroma. However, clinical translation has been limited by the poor reliability of sensor data in real-world conditions. Chemical sensor arrays are highly susceptible to confounding factors such as humidity, temperature gradients, ambient air dilution, and inter-individual variability which distort the VOC scentprint and can degrade classification accuracy of the Al algorithm trained to detect and classify the

[0009] 1 #18994037vlSTR-027W001

[0010] scentprints in health condition specific categories. Among these, humidity is dominant in breath analysis because exhaled samples are naturally near saturation (»95-99% RH at 37 °C). Water molecules compete for adsorption sites, shift surface charge, and alter the dielectric environment of polymer and metal-oxide films, leading to baseline drift, hysteresis, and non-linear responses. The combination of high humidity and chemically complex breath matrices makes it difficult to separate disease-specific VOC temporal response from confounding background, such as physiological noise (defined as exhaled breath humidity, exhalation velocity, exogenous constituents of breath that are not biomarkers for a disease condition).

[0011] [4] Addressing these issues requires a multipronged strategy that prioritizes sensorlevel robustness and measurement discipline: (i) sensing materials with intrinsic hydrophobicity or selective permeability; (ii) coatings and architectures that mitigate water uptake; (iii) controlled sampling with preconditioning (dew-point control and humidityfiltering materials); (iv) on-board environmental metrology (dedicated humidity and temperature channels) with baseline management: and (v) data-driven compensation suited for clinical workflows. Historically, humidity compensation has progressed from purely hardware solutions (preconditioning chambers, flow / temperature control) to hybrid sensor-algorithmic approaches that combine reference channels, signal normalization, and modest regression models for temperature-humidity correction. For example, rapid detection systems using baseline manipulation or orthogonal signal decomposition have been proposed to mitigate humidity drift in MOS arrays, improving VOC selectivity in breath samples.

[0012] Algorithmic interference suppression, such as temperature-humidity compensation via regression models, has suppressed environmental noise in e-noses for aroma quality¬ assessment, but these often require large calibration datasets unavailable in scarce breath enose dataset scenarios. In breath analysis, studies on exhaled VOCs for disease detection (e.g., lung cancer) highlight humidity's role in masking biomarkers, with compensation via moisture filters or signal normalization achieving partial success but failing under variable humidity levels. These methods can substantially reduce environmental variance, but performance can still degrade under rapidly changing humidity, across devices, or when calibration data are scarce.

[0013] [5] Transitioning to machine learning, representation learning methods have been applied to e-nose data to leam robust features invariant to confounders, bridging sensor materials and data processing. Domain adversarial networks, inspired by domain adaptation, train models to minimize confounder influence by aligning distributions across humidity 2 #18994037vlSTR-027W001

[0014] domains, as seen in breath VOC classification where adversarial losses suppress humidity-induced variance. Variational autoencoders (VAEs) enforce probabilistic disentanglement of latent factors, enabling separation of VOC signals from environmental noise in sensor timeseries data. However, these methods often assume large datasets for variance capture, limiting applicability to scarce breath aroma data where overfitting exacerbates confounder entanglement.

[0015] [6] Disentangled representation learning has emerged as a promising extension for explicit confounder isolation in sensor applications, though challenges persist in non-linear, materials-specific interactions. A similarity loss-based framework for disentangling factors in image data is known, but its reliance on pairwise constraints fails to enforce semantic separation in sensor signals with humidity confounders. due to lack of inductive biases in non-linear spaces. Mutual information maximization for paired data have been known, yet without adversarial mechanisms, it struggles with entangled confounders like humidity' that non-linearly interact with VOC responses. DRNET is used for video disentanglement using adversarial training to separate content from pose, but its focus on temporal factors overlooks materials-related confounders such as sensor drift in e-noses. Orthogonality constraints in Vector-Decomposed Disentanglement (VDD) to split domain-invariant and domain-specific representations, are effective for domain shifts but insufficient for humidity 's subtle, nonlinear effects on sensor materials without adversarial debiasing. Disentangled Feature Representation (DFR) for decoupling class-specific features from variations, showing promise in few-shot tasks but assuming separable confounders is known, but it fails in breath data where humidity tightly entangles with VOCs.

[0016] [7] Recent sur eys underscore these limitations in olfaction contexts, noting the need for confounder-robust methods in scarce data regimes. For breath analysis, disentangled adversarial autoencoders have been explored for subject-invariant features in physiological signals, but few target measurable confounders such as humidity in aroma sensors. Invariant Risk Minimization (IRM) and Domain Separation Networks focus on domain shifts, while Variational Fair Autoencoders address fairness, yet none fully disentangle supervised confounders in materials-constrained e-nose data.

[0017] SUMMARY

[0018] [8] The described techniques relate to improved methods, systems, devices, and apparatuses that support techniques for disentangling confounders from volatile organic compound data with adversarial learning. Some implementations may provide a confounder- 3 #18994037vlSTR-027W001

[0019] invariant representation learning framework designed to address the challenges posed by humidity interference in electronic nose systems. The framework may employ a disentangled autoencoder architecture that separates task-relevant volatile organic compound features from confounder-related signals. The encoder component of the system may generate two distinct latent spaces: one dedicated to task-relevant features and another capturing confounder-related information, such as humidity. This separation may be achieved through adversarial training mechanisms, where the encoder may be trained to systematically remove humidity-related features from the task-relevant latent space while preserving them in the confounder-specific latent space.

[0020] [9] Some implementations may include, but not limited to disentangling confounders, such as, velocity of introduction of analyte, exhalation velocity, exhalation pressure, presence of inorganic confounders, such as CO2 from exhaled breath, from the exhaled breath based biomarker.

[0021]

[0010] The decoder may reconstruct the original sensor signal from both latent spaces, ensuring the model learns a complete representation of the input data. Adversarial predictors may be integrated into the framework to isolate and neutralize humidity effects. These predictors may attempt to extract humidity information from the task-relevant latent space, forcing the encoder to refine its representation and achieve confounder invariance. By purging humidity-related distortions, the framework may enable robust classification of volatile organic compounds, even in high-humidity and data-scarce conditions. Some implementations may significantly improve the reliability and generalizability of electronic nose systems, making them suitable for clinical applications such as breath-based diagnostics.

[0022]

[0011] A method for disentangling confounders from volatile organic compound (VOC) data with adversarial learning is described. The method may include receiving sensor data that may include VOC signals and confounder-related signals. The method may include encoding the sensor data into a task-relevant latent representation and a confounder-related latent representation, wherein the encoding may be performed by an encoder that may be configured to separate the VOC signals from the confounder-related signals. The method may include adversarially training a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation. The method may include reconstructing the sensor data from the task-relevant latent representation and the confounder-related latent representation, wherein the reconstruction may be performed by a

[0023] 4 #18994037vlSTR-027W001

[0024] decoder that may be configured to preserve the VOC signals and isolate the confounder-related signals.

[0025]

[0012] A system configured for disentangling confounders from volatile organic compound data with adversarial learning is described. The system may include a processor and memory7coupled with the processor. The system may include instructions stored in the memory and executable by the processor to cause the system to receive sensor data including volatile organic compound signals and confounder-related signals. The system may encode the sensor data into a task-relevant latent representation and a confounder-related latent representation, the encoding being performed by an encoder configured to separate the volatile organic compound signals from the confounder-related signals. The system may adversarially train a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation. The system may reconstruct the sensor data from the task-relevant latent representation and the confounder-related latent representation, the reconstruction being performed by a decoder configured to preserve the volatile organic compound signals and isolate the confounder-related signals.

[0026]

[0013] Another system for disentangling confounders from volatile organic compound data with adversarial learning is described. The system may include means for receiving sensor data including volatile organic compound signals and confounder-related signals. The system may include means for encoding the sensor data into a task-relevant latent representation and a confounder-related latent representation, wherein the encoding may be performed by an encoder configured to separate the volatile organic compound signals from the confounder-related signals. The system may include means for adversarially training a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation. The system may include means for reconstructing the sensor data from the task-relevant latent representation and the confounder-related latent representation, wherein the reconstruction may be performed by a decoder configured to preserve the volatile organic compound signals and isolate the confounder-related signals.

[0027]

[0014] A non-transitory computer-readable medium storing code for disentangling confounders from volatile organic compound data is described. The code may include instructions executable by a processor to receive sensor data including volatile organic compound signals and confounder-related signals. The code may include instructions executable by a processor to encode the sensor data into a task-relevant latent representation 5 #18994037vlSTR-027W001

[0028] and a confounder-related latent representation, wherein the encoding may be performed by an encoder configured to separate the volatile organic compound signals from the confounder-related signals. The code may include instructions executable by a processor to adversarially train a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation. The code may include instructions executable by a processor to reconstruct the sensor data from the task-relevant latent representation and the confounder-related latent representation, wherein the reconstruction may be performed by a decoder configured to preserve the volatile organic compound signals and isolate the confounder-related signals.

[0029]

[0015] Some examples of the method, systems, and non-transitory computer-readable medium described herein may further include operations, features, means, or instructions for normalizing the sensor data by adjusting the volatile organic compound signals and the confounder-related signals based on a reference baseline derived from ambient sampling prior to encoding the sensor data.

[0030]

[0016] Some examples of the method, systems, and non-transitory computer-readable medium described herein may further include operations, features, means, or instructions for truncating the sensor data to exclude uninformative temporal regions by retaining a fixed window encompassing baseline, exposure, and recovery phases prior to encoding the sensor data.

[0031]

[0017] Some examples of the method, systems, and non-transitory computer-readable medium described herein may further include operations, features, means, or instructions for applying zero-padding to the sensor data sequences shorter than a predefined length to maintain uniform tensor dimensions for batch processing prior to encoding the sensor data.

[0032]

[0018] Some examples of the method, systems, and non-transitory computer-readable medium described herein may further include operations, features, means, or instructions for weighting the task-relevant latent representation during training to emphasize minority class signals in imbalanced datasets while suppressing overrepresentation of majority class signals.

[0033]

[0019] Some examples of the method, systems, and non-transitory computer-readable medium described herein may further include operations, features, means, or instructions for training the decoder to reconstruct the sensor data by combining the task-relevant latent representation and the confounder-related latent representation while preserving temporal causality in the reconstructed signals.

[0034] 6 #18994037vlSTR-027W001

[0035]

[0020] In some examples of the method, systems, and non-transitory computer-readable medium described herein, the encoder may be configured to apply temporal convolutional layers to the sensor data to extract sequential features while reducing noise from the confounder-related signals.

[0036]

[0021] In some examples of the method, systems, and non-transitory computer-readable medium described herein, the confounder predictor may be trained to identify confounder-related signals by minimizing a loss function specific to the confounder-related latent representation.

[0037]

[0022] In some examples of the method, systems, and non-transitory computer-readable medium described herein, the task-relevant latent representation may be configured to retain volatile organic compound signals by suppressing features correlated with the confounder-related latent representation.

[0038]

[0023] In some examples of the method, systems, and non-transitory computer-readable medium described herein, the reconstruction performed by the decoder may include aligning the temporal structure of the reconstructed sensor data with the original sensor data.

[0039]

[0024] In some examples of the method, systems, and non-transitory computer-readable medium described herein, the adversarial training may include dynamically adjusting the learning rate of the confounder predictor to enhance separation between the task-relevant and confounder-related latent representations.

[0040]

[0025] In some examples of the method, systems, and non-transitory computer-readable medium described herein, the encoder may be configured to generate the task-relevant latent representation and the confounder-related latent representation by applying distinct activation functions to separate task-relevant and confounder-related features.

[0041] BRIEF DESCRIPTION OF THE DRAWINGS

[0042]

[0026] FIG. 1 illustrates an aroma sensor array, also known as aroma chip, that supports disentangling confounders from volatile organic compound data with adversarial learning in accordance with aspects of the present disclosure.

[0043]

[0027] FIG. 2 shows an aroma sensor setup which supports techniques for disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure.

[0044]

[0028] FIG. 3 shows a breathalyzer device setup which supports techniques for disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure.

[0045] 7 #18994037vlSTR-027W001

[0046]

[0029] FIG. 4 shows an architecture comparison diagram which supports techniques for disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure.

[0047]

[0030] FIG. 5 shows a block diagram of an apparatus that supports disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure.

[0048]

[0031] FIG. 6 shows a block diagram of a confounder disentanglement component that supports disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure.

[0049]

[0032] FIG. 7 shows a diagram of a system including a device that supports disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure.

[0050]

[0033] FIG. 8 shows a flowchart illustrating a method that supports disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure.

[0051] DETAILED DESCRIPTION

[0052]

[0034] Methods, systems, devices, and apparatuses that support techniques for disentangling confounders from volatile organic compound data with adversarial learning are disclosed. In some examples, electronic nose systems may face significant challenges in reliably detecting volatile organic compounds in environments with high humidity. Humidity, a dominant confounder in breath analysis, may compete for adsorption sites on sensor materials, alter surface charge, and introduce non-linear distortions in sensor responses. These effects may obscure subtle volatile organic compound signals, degrade classification accuracy, and hinder the development of generalizable machine learning models. Existing approaches, such as hardware-based preconditioning and regression-based compensation, may provide partial solutions but may be inadequate in scenarios with limited data or rapidly- changing environmental conditions. Furthermore, current representation learning methods may often fail to disentangle confounder effects from task-relevant features, leading to overfitting and reduced robustness in electronic nose applications. This problem may be particularly acute in clinical settings, where reliable detection of volatile organic compounds may be critical for non-invasive diagnostics.

[0053]

[0035] In some implementations, a system may be designed to analyze data from electronic nose devices, which may mimic the human sense of smell to detect patterns of 8 #18994037vlSTR-027W001

[0054] volatile organic compounds. These implementations may include a framework for learning data representations that may remain unaffected by external factors, such as humidity, aroma introduction velocity, pressure, physiological features pertaining to demographics, which may interfere with the detection of these compounds, which may be biomarkers for a health condition. The framework may be particularly suited for environments with high humidity, such as breath analysis, and may be optimized to function effectively even when limited training data may be available. To achieve this, the system may employ advanced techniques to disentangle relevant information from confounding variables, ensuring that the analysis may remain robust under challenging conditions. This approach may enhance the reliability and reproducibility of electronic nose devices, making them suitable for applications such as point-of-care diagnostics and long-term monitoring.

[0055]

[0036] In some implementations, the system may include a disentangled autoencoder, which may consist of three main components: an encoder, a decoder, and predictors. The encoder may process input data from sensors and separate it into two distinct representations. One representation may focus on identifying the compounds being analyzed, while the other may isolate information related to external factors, such as humidity. The decoder may then reconstruct the original input data using both representations, ensuring that the system may leam a complete and accurate representation of the data. Additionally, the predictors may perform specific tasks: one may classify compounds using the purified representation generated by the encoder, while the other may attempt to extract information about external factors from the same representation. This setup may allow the encoder to systematically remove information related to external factors from the data used for classification, ensuring that the system may focus on the relevant chemical signatures.

[0056]

[0037] In some implementations, the training process may involve adversarial dynamics, where the encoder and the predictor for external factors may be trained in opposition to each other. The encoder may aim to create a representation of the data that may be free from external factors, while the predictor may attempt to extract information about these factors from the same representation. This adversarial interaction may ensure that the encoder effectively removes information related to external factors, such as humidity, from the data used for identifying compounds. To optimize this process, the system may use a combination of objectives, including accurate reconstruction of input data by the decoder, effective classification of compounds by the predictors, and systematic removal of external factor information by the encoder. Adjustable parameters may be used to balance these objectives, ensuring that the system may achieve optimal performance across all tasks.

[0057] 9 #18994037vlSTR-027W001

[0058]

[0038] In some implementations, the encoder may include specialized layers designed to process time-series data, which may be collected over time from sensors. These layers may generate two separate representations: one optimized for identifying compounds and another optimized for isolating information related to external factors. To enhance stability and performance during training, the encoder may also include additional components, such as normalization layers and activation functions. The decoder, in turn, may use layers that reverse the processing performed by the encoder to reconstruct the original input data. This reconstruction process may ensure that the representations generated by the encoder retain sufficient information to accurately represent the data collected by sensors, even after disentangling external factors.

[0059]

[0039] In some implementations, predictors may play a critical role in the system by performing specific tasks related to classification and external factor analysis. One predictor may classify compounds using the purified representation generated by the encoder, while another predictor may attempt to extract information about external factors, such as humidity, from the same representation. The predictor for external factors may include additional layers to reconstruct signals related to these factors, ensuring that the system may accurately isolate and neutralize their effects. By systematically purging the representation used for classification of information related to external factors, the system may achieve high accuracy in identifying compounds while retaining detailed information about external conditions in a separate representation. This dual capability may make the system particularly effective in environments with complex and variable conditions.

[0060]

[0040] In some implementations, the training process may involve simultaneous optimization of the encoder, decoder, and predictors, with metrics used to monitor the adversarial dynamics during training. For example, the system may measure the accuracy of reconstructing signals related to external factors from the representations generated by the encoder. Successful isolation of external factors may be indicated by low accuracy in reconstructing these signals from the representation used for classification and high accuracy in reconstructing them from the representation dedicated to external factors. To further enhance performance, the system may preprocess data collected by sensors before training. This preprocessing may include normalizing sensor responses to account for variability and drift, truncating sequences to capture relevant dynamics while removing uninformative data, and padding sequences to ensure uniform dimensions for processing.

[0061]

[0041] In some implementations, the system may be validated using datasets that may involve analyzing compounds in various scenarios, such as acetone in controlled

[0062] 10 #18994037vlSTR-027W001

[0063] environments, breath samples from individuals on specific diets, and breath samples before and after ingesting certain substances. These datasets may present challenges such as high humidity, detecting compounds in trace amounts, and imbalanced data, and the system may be tested for its ability to handle these conditions. Additionally, studies may be conducted to evaluate the contribution of individual components within the system, confirming that the reconstruction process and adversarial dynamics may be essential for achieving the separation of information related to external factors from information relevant to identifying compounds. These studies may provide insights into the system’s design and highlight areas for further optimization.

[0064]

[0042] In some implementations, the system may generalize across different types of compounds, including those that may interact differently with water, such as hydrophilic and hydrophobic compounds. The system may be designed to isolate information related to external factors, such as humidity, regardless of the specific chemical properties of the compounds being analyzed. Theoretical guarantees may support this capability by using measures of shared information between variables to quantify the separation and ensure that the representations generated by the system may be identifiable and predictive. Furthermore, the system may include a bound that may assess its ability to generalize to new data, particularly in scenarios where limited data may be available. This bound may highlight the system’s efficiency in reducing overfitting and improving robustness in small datasets.

[0065]

[0043] In some implementations, the system may be particularly suited for analyzing breath samples in environments with high humidity and detecting compounds in trace amounts. By enhancing reproducibility and reliability, the system may improve the performance of electronic nose devices in various applications, such as point-of-care diagnostics and monitoring over time. The system may also be optimized for isolating a single external factor, such as humidity, and may be extended to handle multiple factors, such as temperature and sensor drift. These extensions may improve the system’s ability’ to analyze data affected by various conditions, making it more versatile and robust. Additionally, the system may be integrated into electronic nose devices that may use sensor arrays to detect compounds, with components for measuring external factors, such as humidity and temperature, to provide real-time data for isolating these factors.

[0066]

[0044] In some implementations, standardized protocols may be used for collecting data, which may include phases for sampling ambient air, sampling compounds, and recovering sensors. These protocols may ensure consistency across experiments, enabling reliable comparisons and reproducible results. To further enhance performance, the system may 11 #18994037vlSTR-027W001

[0067] optimize adjustable parameters using methods that may explore different configurations to achieve balanced performance across objectives, such as reconstructing data, classifying compounds, and isolating external factors. By addressing these challenges, the system may provide a robust and efficient solution for analyzing complex datasets in diverse environments.

[0068]

[0045] Aspects of the subject matter described in this disclosure can be implemented to realize one or more of the following potential advantages. The described techniques may be implemented to support improved robustness in environments where external factors, such as humidify, may interfere with sensor data, ensuring that the analysis may remain reliable under variable conditions. The system may enable the detection of volatile organic compounds in scenarios where data availability may be limited, allowing for effective performance in data-scarce applications. By leveraging disentangled representations, the system may isolate confounding variables from task-relevant features, which may enhance the accuracy of classification tasks in complex datasets. The framework may be adaptable to different sensor technologies, making it suitable for diverse electronic nose configurations that may require compensation for environmental influences. The described methods may facilitate the development of diagnostic tools that may operate effectively in real-world settings, where external factors may otherwise compromise the reliability' of results.

[0069]

[0046] Aspects of the disclosure are initially described in the context of networked computing systems. Aspects of the disclosure are additionally illustrated by and described with reference to example implementations. Aspects of the disclosure are further illustrated by and described with reference to apparatus diagrams, system diagrams, and flowcharts that relate to disentangling confounders from volatile organic compound data with adversarial learning.

[0070]

[0047] FIG. 1 illustrates an example of an aroma sensor array that is used for implementing the disentanglement. The aroma sensor array 101 comprises a plurality of sensing elements, 104 with interdigitated electrodes. A serial communication bus 102, is provided with the aroma chip, along with the confounder ground truth detection sensor, 103.

[0071]

[0048] FIG. 2 shows aroma sensor setup 200 which supports techniques for disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure. As depicted in FIG. 2, the aroma sensor setup 200 may include one or more of a chemiresistive sensor array chip 205, an aroma sensing headspace 201. an aroma sample inlet 202, an aroma sample outlet 203, a piezoelectric pump 204, and / or other components.

[0072] 12 #18994037vlSTR-027W001

[0073]

[0049] The chemiresistive sensor array chip 205 may include an array of sensing elements designed to detect volatile organic compounds. The chemiresistive sensor array chip 205 may feature thin films composed of polymer-carbon black nanocomposites, which may exhibit changes in electrical resistance upon exposure to volatile organic compounds. The chemiresistive sensor array chip 205 may generate a 32-dimensional time-series data output, which may represent the unique scentprint of the aroma sample. The chemiresistive sensor array chip 205 may capture aroma samples under controlled conditions through the headspace and piezoelectric pump. In some implementations, alternative sensing materials, such as metal-oxide semiconductors or graphene-based sensors, may be used in place of polymercarbon black nanocomposites.

[0074]

[0050] The interdigitated electrode may represent a patterned structure that facilitates electrical measurements within the sensor array. The interdigitated electrode may include ENIG-plated gold electrodes patterned on copper traces over a Rogers substrate. The interdigitated electrode may feature electrode fingers with a width of 100 pm, an interelectrode gap of 100 pm, and an electrode height of approximately 12 pm. The interdigitated electrode may be integral to the chemiresistive sensor array chip, enabling precise resistance measurements during aroma sampling. In some implementations, alternative electrode materials, such as platinum or silver, may be used to achieve similar functionality.

[0075]

[0051] The aroma sensing headspace 201 may provide an enclosed environment for capturing aroma samples during analysis. The aroma sensing headspace 201 may be designed to hold the chemiresistive sensor array chip 205 and may include a pump for active aroma sampling. The aroma sensing headspace 201may maintain controlled conditions to prevent external contamination of the aroma sample. The aroma sensing headspace 201 may connect to the aroma sample inlet 202 and the aroma sample outlet 203 to facilitate the flow of aroma samples. In some implementations, the aroma sensing headspace 201 may be constructed from materials such as glass or stainless steel to ensure chemical inertness.

[0076]

[0052] The aroma sample inlet 202 may include a pathway for introducing aroma samples into the sensing headspace. The aroma sample inlet 202 may be connected to a PTFE tube that channels aroma samples from external sources, such as vials or breath sampling modules, into the aroma sensing headspace 201. The aroma sample inlet 202 may be designed to prevent backflow and contamination during aroma sampling. The aroma sample inlet 202 may work in conjunction with the piezoelectric pump 204 to regulate the flow rate of incoming aroma samples. In some implementations, the aroma sample inlet 202 may feature additional filters to remove particulate matter or microbial contaminants.

[0077] 13 #18994037vlSTR-027W001

[0078]

[0053] The aroma sample outlet 203 may represent an exit channel for removing aroma samples from the sensing headspace. The aroma sample outlet 203 may be connected to a PTFE tube that directs the aroma sample away from the aroma sensing headspace 201 after analysis. The aroma sample outlet 203 may ensure the removal of residual volatile organic compounds to prepare the sensor array for subsequent aroma sampling. The aroma sample outlet 203 may operate in coordination with the piezoelectric pump 204 to maintain consistent airflow during the recovery phase. In some implementations, the aroma sample outlet 203 may include additional components, such as valves or filters, to control the exhaust flow.

[0079]

[0054] The piezoelectric pump 204 may provide controlled airflow to support aroma sampling within the setup. The piezoelectric pump 204 may operate at a flow rate of 10 mL / min to ensure consistent sampling conditions during all phases of the aroma digitization protocol. The piezoelectric pump 204 may be integrated into the aroma sensing headspace 201to facilitate the movement of aroma samples through the aroma sample inlet 202 and aroma sample outlet 203. The piezoelectric pump 204 may be used to flush the sensor array during the recovery phase to remove residual volatile organic compounds. In some implementations, alternative pump mechanisms, such as diaphragm pumps or peristaltic pumps, may be used to achieve similar airflow control.

[0080]

[0055] In some implementations, the chemiresistive sensor array chip 205 may be positioned within the aroma sensing headspace 201 to interact directly with the incoming aroma sample. The interdigitated electrodes on the chip may facilitate the detection of resistance changes in the chemiresistive thin films, which may correspond to the presence of volatile organic compounds (VOCs). The aroma sample inlet 202 may direct the sample into the headspace 201. where the controlled environment may allow the sensor array chip 205 to capture the chemical profile of the sample.

[0081]

[0056] In some implementations, the piezoelectric pump 204 may regulate the airflow through the aroma sensing headspace 201, ensuring a consistent flow rate for sampling. The aroma sample outlet 203 may expel the processed sample, maintaining the internal conditions of the headspace 201. The integration of the chemiresistive sensor array chip 205 with the piezoelectric pump 204 and the aroma sensing headspace 201 may allow for the sequential capture, analysis, and expulsion of aroma samples, wi th the interdigitated electrodes playing a key role in translating chemical interactions into measurable electrical signals.

[0082]

[0057] FIG. 3 shows breathalyzer device setup 300 which supports techniques for disentangling confounders from volatile organic compound data with adversarial learning in 14 #18994037vlSTR-027W001

[0083] accordance with various aspects of the present disclosure. As depicted in FIG. 3, the breathalyzer device setup 300 may include one or more of a noze aroma chip integrated into the breathalyzer, a breath sampling buffer 301 of the breathalyzer and a breathalyzer mouthpiece 302, and / or other components.

[0084]

[0058] The noze aroma chip may include an array of chemiresistive sensing elements designed to detect volatile organic compounds in exhaled breath. The noze aroma chip may feature 32 distinct chemiresistive sensing elements, each fabricated with polymer-carbon black nanocomposites approximately 1 pm thick. The substrate of the noze aroma chip may be constructed using Electroless Nickel Immersion Gold-plated interdigitated electrodes patterned on copper traces over a 1 mm Rogers substrate. The noze aroma chip may measure resistance changes at a sampling rate of 1 Hz, generating a 32-dimensional time-series that may represent the unique scentprint of the aroma sample. In some implementations, the noze aroma chip may be integrated into the breathalyzer to digitize exhaled breath samples for analysis. The noze aroma chip may be enclosed in an aroma sensing headspace 201, which may be filled through a piezoelectric pump 204 for active aroma sampling.

[0085]

[0059] The breath sampling buffer 301 may represent an internal chamber configured to collect and temporarily store the alveolar portion of exhaled breath for analysis. The breath sampling buffer 301 may include a capnography valve that may allow only the alveolar portion of exhaled breath, enriched with metabolic volatile organic compounds, to enter the buffer chamber. The breath sampling buffer 301 may be connected to the noze aroma chip, which may digitize the collected breath sample into a scentprint. In some implementations, the breath sampling buffer 301 may be part of the breathalyzer setup, which may include a pump to control airflow during sampling phases. The breath sampling buffer 301 may be designed to work with a detachable breathalyzer mouthpiece 302 to ensure safe and controlled breath sampling.

[0086]

[0060] The breathalyzer mouthpiece 302 may provide a detachable interface equipped with a microbial fdter, humidity fdter, and backflow prevention valve for safe and controlled breath sampling. The breathalyzer mouthpiece 302 may include a single-patient-use design to ensure hygiene and safety during breath sampling. The microbial filter of the breathalyzer mouthpiece 302 may prevent contamination, while the humidity filter may reduce the impact of high humidity levels in exhaled breath samples. The backflow prevention valve of the breathalyzer mouthpiece 302 may ensure that exhaled breath flows in one direction into the breath sampling buffer 301. In some implementations, the breathalyzer mouthpiece 302 may be connected to the breath sampling buffer 301, which may collect the alveolar portion of 15 #18994037vlSTR-027W001

[0087] exhaled breath for analysis. The breathalyzer mouthpiece 302 may be compatible with the DiagNoze breathalyzer setup, which may include a pump to control airflow during sampling phases.

[0088]

[0061] The aroma sample inlet 202 may be configured to introduce aroma samples into the aroma sensing headspace 201for analysis. The aroma sample inlet 202 may be connected a plurality of different introduction apparatus, to introduce the aroma into the headspace of the aroma sensor array

[0089]

[0062] The aroma sample outlet 203 may be configured to remove aroma samples from the aroma sensing headspace 201 after analysis. The aroma sample outlet 203 may be connected to the ambient air to flush away volatile organic compounds from the chemiresistive sensor array chip 205 during the sensor recovery phase. The aroma sample outlet 203 may work in conjunction with the piezoelectric pump 204 to ensure controlled airflow during the recovery phase. In some implementations, the aroma sample outlet 203 may be part of the vial-based aroma sampler setup, which may include a standardized aroma digitization protocol consisting of ambient sampling, aroma sampling, and sensor recovery phases.

[0090]

[0063] The aroma sensing headspace 201 may be configured to hold the noze aroma chip for aroma sample analysis. The aroma sensing headspace 201 may be enclosed in a specially designed setup that may include a piezoelectric pump 204 to control incoming airflow. The aroma sensing headspace 201 may be connected to the aroma sample inlet 202 and aroma sample outlet 203 to facilitate the transfer of aroma samples during sampling and recovery phases. In some implementations, the aroma sensing headspace 201may be part of the vial-based aroma sampler setup, which may include a pump operating at a flow rate of 10 mL / min to transfer aroma samples into the headspace for analysis.

[0091]

[0064] The chemiresistive sensor array chip 205 may include an array of chemiresistive sensing elements designed to detect volatile organic compounds in aroma samples. The chemiresistive sensor array chip 205 may feature 32 distinct sensing elements, each fabricated with polymer-carbon black nanocomposites approximately 1 pm thick. The substrate of the chemiresistive sensor array chip 205 may be constructed using Electroless Nickel Immersion Gold-plated interdigitated electrodes patterned on copper traces over a 1 mm Rogers substrate. The chemiresistive sensor array chip 205 may measure resistance changes at a sampling rate of 1 Hz, generating a 32-dimensional time-series that may represent the unique scentprint of the aroma sample. In some implementations, the

[0092] 16 #18994037vlSTR-027W001

[0093] chemiresistive sensor array chip 205 may be integrated into the aroma sensing headspace 201, which may be filled through a piezoelectric pump 204 for active aroma sampling.

[0094]

[0065] The interdigitated electrode may be part of the chemiresistive sensor array chip 205 and may be configured to detect resistance changes upon volatile organic compound adsorption. The interdigitated electrode may include a width of 100 pm, an inter-electrode gap of 100 pm, and an electrode height of approximately 12 pm. The interdigitated electrode may be plated with Electroless Nickel Immersion Gold to enhance conductivity and durability. In some implementations, the interdigitated electrode may be part of the chemiresistive sensor array chip 205, which may be integrated into the aroma sensing headspace 201for aroma sample analysis.

[0095]

[0066] In some implementations, the noze aroma chip may be integrated into the breathalyzer housing to interact directly with the breath sampling buffer 301, which may temporarily hold the exhaled breath sample. The breath sampling buffer 301 may be positioned to channel the alveolar portion of the breath sample toward the noze aroma chip for digitization into a scentprint. The breathalyzer mouthpiece 302 may be detachable and may include a microbial filter, humidity filter, and backflow prevention valve, which may collectively regulate the flow and composition of the breath sample entering the buffer 301.

[0096]

[0067] In some implementations, the pump mechanism within the breathalyzer may control airflow through the breath sampling buffer 301, ensuring consistent delivery of the sample to the noze aroma chip. The noze aroma chip may feature an array of chemiresistive sensing elements that may detect VOCs in the breath sample, generating a multidimensional time-series representation. The breathalyzer may have capnography based implementations to isolate the alveolar portion of the breath sample, which may contain metabolic VOCs, before directing it into the buffer 301 for processing.

[0097]

[0068] FIG. 4 shows architecture comparison diagram 400 which supports techniques for disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure. As depicted in FIG. 4, the architecture comparison diagram 400 may include one or more of an input data 401, an encoder 405, a latent space , a decoder 406, a task classifier 408, a confounder predictor 409, a prediction 404, and / or other components.

[0098]

[0069] The input data 401 may include time-series signals representing volatile organic compound (VOC) sensor responses. The input data 401 may originate from chemiresistive sensing elements that detect changes in electrical resistance upon VOC adsorption. The input data 401 may represent a unique “scentprint” of the aroma sample, capturing temporal 17 #18994037vlSTR-027W001

[0099] variations in sensor responses. The input data 401 may be normalized to account for baseline drift and inter-sensor variability. In some implementations, the input data 401 may include signals from other types of sensors, such as piezoelectric or optical sensors.

[0100]

[0070] The encoder 405 may process the input data 401 through a series of layers to generate latent representations. The encoder 405 may include temporal convolutional layers that extract features from the input data 401. The encoder 405 may apply batch normalization and activation functions, such as LeakyReLU, to stabilize training and enhance feature extraction. The encoder 405 may generate two distinct latent spaces: one for task-relevant features and another for confounder-related features. In some implementations, the encoder 405 may be configured to process input data from different sensor modalities.

[0101]

[0071] The latent space may represent a disentangled structure comprising task-relevant and confounder-related features. The latent space may include a task-relevant subspace that captures VOC -specific information and a confounder-related subspace that isolates humidity-related variations. The latent space may be structured to ensure that task-relevant features are invariant to confounder effects. The latent space may be parameterized to allow flexible dimensionality for both subspaces. In some implementations, the latent space may be used to represent other confounding factors, such as temperature or sensor drift.

[0102]

[0072] The decoder 406 may reconstruct the input data 401 from the latent space . The decoder 406 may include transposed convolutional layers that map latent representations back to the original input space. The decoder 406 may use learned weights to reconstruct temporal patterns in the input data 401. The decoder 406 may ensure that both task-relevant and confounder-related features are accurately represented in the reconstruction. In some implementations, the decoder 406 may be adapted to reconstruct data from other sensor types or configurations.

[0103]

[0073] The task classifier 408 may determine task-specific predictions based on the taskrelevant features in the latent space . The task classifier 408 may include fully connected layers that process the task-relevant subspace of the latent space . The task classifier 408 may apply dropout to reduce overfitting during training. The task classifier 408 may output predictions corresponding to VOC concentrations or classifications. In some implementations, the task classifier 408 may be designed to predict other task-specific labels, such as disease states or aroma profiles.

[0104]

[0074] The confounder predictor 409 may attempt to determine confounder-related information from the latent space . The confounder predictor 409 may include fully connected layers that process the confounder-related subspace of the latent space . The 18 #18994037vlSTR-027W001

[0105] confounder predictor 409 may be trained adversarially to extract humidity -related features from the latent space . The confounder predictor 409 may output predictions corresponding to humidity levels in the aroma sample. In some implementations, the confounder predictor 409 may be configured to predict other environmental factors, such as temperature or air pressure.

[0106]

[0075] The prediction 404 may represent the output generated by the task classifier 408. The prediction 404 may include task-specific labels, such as VOC concentrations or aroma classifications. The prediction 404 may be generated based on the purified task-relevant features in the latent space . The prediction 404 may be evaluated using metrics such as Flscore, precision, recall, and AUC. In some implementations, the prediction 404 may include additional outputs, such as confidence scores or probabilistic estimates.

[0107]

[0076] In some implementations, the input data 401 may pass through the encoder 405, which may consist of temporal convolutional layers configured to extract features from the sensor signals. The encoder 405 may generate two distinct latent spaces, the task-relevant latent space and the confounder latent space , which may separately capture VOC-specific features and humidity-related information. The decoder 406 may reconstruct the original input signal by combining information from both latent spaces , ensuring that the reconstructed signal may retain fidelity to the input data 401.

[0108]

[0077] In some implementations, the task classifier 408 may operate on the task-relevant latent space to determine the VOC classification label, while the confounder predictor 409 may attempt to extract humidity-related features from the same latent space . The adversarial interaction between the encoder 405 and the confounder predictor 409 may guide the encoder 405 to systematically remove humidity -related information from the task-relevant latent space . The prediction 404 may represent the output of the task classifier 408, which may reflect the classification result based on the purified task-relevant representation. 407, 410 and 411 represent the reconstructed latent space, latent space for task and latent space for confounder respectively.

[0109]

[0078] FIG. 5 shows a block diagram 500 of an apparatus 502 that supports disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure. The apparatus 502 may include an input module 504, confounder disentanglement component 506, and an output module 508. The apparatus 502 may also include a processor. Each of these components may be in communication with one another (e.g., via one or more buses). In some cases, the apparatus 502 may be an example of a user terminal, a database server, or a system containing multiple computing devices.

[0110] 19 #18994037vlSTR-027W001

[0111]

[0079] The input module 504 may manage input signals for the apparatus 502. For example, the input module 504 may identify input signals based on an interaction with a modem, a keyboard, a mouse, a touchscreen, or a similar device. These input signals may be associated with user input or processing at other components or devices. In some cases, the input module 504 may utilize an operating system such as iOS®, ANDROID®, MS-DOS®, MS-WINDOWS®, OS / 2®, UNIX®, LINUX®, or another known operating system to handle input signals. The input module 504 may send aspects of these input signals to other components of the apparatus 502 for processing. In some cases, the input module 504 may be a component of an input / output (I / O) controller 706 as described with reference to FIG. 7.

[0112]

[0080] The confounder disentanglement component 506 may include one or more of a sensor data receiving component 510, an encoder configuration component 512, a confounder predictor training component 514, a sensor data reconstruction component 516, and / or other components. The confounder disentanglement component 506 may be an example of aspects of the confounder disentanglement component 602 or 704 described with reference to FIGS.

[0113] 6 and 7.

[0114]

[0081] The sensor data receiving component 510 may be configured as or otherwise support a means for receiving sensor data including volatile organic compound signals and confounder-related signals. The encoder configuration component 512 may be configured as or otherw ise support a means for encoding the sensor data into a task-relevant latent representation and a confounder-related latent representation, wherein the encoder may be configured to separate the volatile organic compound signals from the confounder-related signals. The confounder predictor training component 514 may be configured as or otherwise support a means for adversarially training a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation. The sensor data reconstruction component 516 may be configured as or otherwise support a means for reconstructing the sensor data from the task-relevant latent representation and the confounder-related latent representation, wherein the decoder may be configured to preserve the volatile organic compound signals and isolate the confounder-related signals.

[0115]

[0082] The output module 508 may manage output signals for the apparatus 502. For example, the output module 508 may receive signals from other components of the apparatus 502, such as the confounder disentanglement component 506. and may transmit these signals to other components or devices. In some specific examples, the output module 508 may transmit output signals for display in a user interface, for storage in a database or data store,

[0116] 20 #18994037vlSTR-027W001

[0117] for further processing at a server or server cluster, or for any other processes at any number of devices or systems. In some cases, the output module 508 may be a component of an I / O controller 706 as described with reference to FIG. 7.

[0118]

[0083] FIG. 6 shows a block diagram 600 of a confounder disentanglement component 602 that supports disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure. The confounder disentanglement component 602 may be an example of aspects of a confounder disentanglement component 506, a confounder disentanglement component 704, or both, as described herein. The confounder disentanglement component 602, or various components thereof, may be an example of means for performing various aspects of disentangling confounders from volatile organic compound data with adversarial learning as described herein. For example, the confounder disentanglement component 602 may include one or more of a sensor data receiving component 604, an encoder configuration component 606, a confounder predictor training component 608, a sensor data reconstruction component 610, a data normalization component 612, a temporal truncation component 614, a sequence padding component 616, a latent representation weighting component 618, a decoder training component 620, and / or other components. Each of these components may communicate, directly or indirectly, with one another (e.g., via one or more buses).

[0119]

[0084] The sensor data receiving component 604 may be configured as or otherwise support a means for receiving sensor data including volatile organic compound signals and conform der-related signals. The sensor data receiving component 604 may include an interface that may support real-time data acquisition from chemiresistive sensor arrays. The sensor data receiving component 604 may be designed to handle time-series data streams that may originate from aroma sensor arrays with multiple sensing elements. The sensor data receiving component 604 may incorporate mechanisms to normalize incoming signals, which may account for baseline drift in sensor responses. The sensor data receiving component 604 may support integration with environmental metrology7sensors, which may include humidity and temperature channels for concurrent data collection. The sensor data receiving component 604 may process signals at a sampling rate that may align with the temporal resolution of the connected sensor array. The sensor data receiving component 604 may accommodate data from diverse sensor materials, which may include polymer-carbon black nanocomposites or metal-oxide semiconductors. The sensor data receiving component 604 may be compatible with headspace sampling setups that may involve controlled airflow

[0120] 21 #18994037vlSTR-027W001

[0121] mechanisms. The sensor data receiving component 604 may support preprocessing steps that may include ambient normalization and temporal sequence truncation.

[0122]

[0085] The encoder configuration component 606 may be configured as or otherwise support a means for encoding the sensor data into a task-relevant latent representation and a confounder-related latent representation. In some implementations, the encoder configuration component 606 may include temporal convolutional layers that may process time-series data to extract relevant features. The encoder configuration component 606 may accommodate latent spaces with adjustable dimensionality to align with the complexity of the sensor data. The encoder configuration component 606 may support disentanglement mechanisms that may isolate humidity-related signals from volatile organic compound signals.

[0123]

[0086] The encoding may be performed by an encoder that may be configured to separate the volatile organic compound signals from the confounder-related signals. In some implementations, the encoder may incorporate adversarial training techniques that may refine the separation of task-relevant and confounder-related features. The encoder may include batch normalization layers that may stabilize the encoding process across varying sensor inputs. The encoder may utilize activation functions, such as LeakyReLU, that may enhance the representation of non-linear sensor data patterns.

[0124]

[0087] The confounder predictor training component 608 may be configured as or otherwise support a means for adversarially training a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation. In some implementations, the confounder predictor training component 608 may incorporate gradient reversal layers that may invert gradients during backpropagation to enforce adversarial dynamics. The confounder predictor training component 608 may include mechanisms to adjust the learning rate dynamically, which may align with the complexity of the confounder-related latent representation. The confounder predictor training component 608 may support training with batch-wise data augmentation techniques that may introduce variability in confounder-related signals to enhance the robustness of the predictor.

[0125]

[0088] The confounder predictor training component 608 may be configured as or otherwise support a means for adversarially training a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation. In some implementations, the confounder predictor training component 608 may utilize loss functions that may penalize the predictor for extracting task-relevant signals 22 #18994037vlSTR-027W001

[0126] from the confounder-related latent representation. The confounder predictor training component 608 may accommodate multi-channel input data that may include humidity and temperature signals alongside volatile organic compound signals. The confounder predictor training component 608 may integrate regularization techniques that may prevent overfitting to specific confounder patterns during training.

[0127]

[0089] The sensor data reconstruction component 610 may be configured as or otherwise support a means for reconstructing the sensor data from the task-relevant latent representation and the confounder-related latent representation. In some implementations, the sensor data reconstruction component 610 may incorporate temporal sequence alignment techniques to ensure that the reconstructed signals align with the original time-series data. The sensor data reconstruction component 610 may accommodate reconstruction processes that may handle varying levels of noise in the input data to maintain fidelity.

[0128]

[0090] The reconstruction may be performed by a decoder that may be configured to preserve the volatile organic compound signals and may isolate the confounder-related signals. In some implementations, the decoder may include transposed convolutional layers that may reconstruct high-dimensional sensor data from the latent representations. The decoder may support mechanisms to adjust reconstruction parameters dynamically based on the characteristics of the input latent representations.

[0129]

[0091] In some examples, the data normalization component 612 may be configured as or otherwise support a means for normalizing the sensor data by adjusting the volatile organic compound signals and the confounder-related signals based on a reference baseline derived from ambient sampling prior to encoding the sensor data. In some implementations, the data normalization component 612 may determine the reference baseline by analyzing ambient air samples collected during a predefined sampling phase. In some implementations, the data normalization component 612 may incorporate algorithms that may account for temporal variations in ambient conditions to refine the reference baseline. In some implementations, the data normalization component 612 may adjust the volatile organic compound signals by applying scaling factors that may align the sensor responses with the determined baseline. In some implementations, the data normalization component 612 may adjust the confounder-related signals by applying offset corrections that may compensate for baseline drift observed during ambient sampling.

[0130]

[0092] In some examples, the temporal truncation component 614 may be configured as or otherwise support a means for truncating the sensor data to exclude uninformative temporal regions by retaining a fixed window that may encompass baseline, exposure, and 23 #18994037vlSTR-027W001

[0131] recovery phases prior to encoding the sensor data. In some implementations, the temporal truncation component 614 may determine the fixed window duration based on the sampling rate of the sensor array to align with the temporal resolution of the data. In some implementations, the temporal truncation component 614 may accommodate variable-length input sequences by applying zero-padding to shorter sequences to maintain consistent dimensions for subsequent processing. In some implementations, the temporal truncation component 614 may incorporate mechanisms to identify and exclude signal artifacts that may occur outside the defined temporal window.

[0132]

[0093] In some examples, the sequence padding component 616 may be configured as or otherwise support a means for applying zero-padding to the sensor data sequences that may be shorter than a predefined length to maintain uniform tensor dimensions for batch processing prior to encoding the sensor data. In some implementations, the sequence padding component 616 may determine the predefined length based on the average duration of sensor data sequences collected during aroma sampling experiments. In some implementations, the sequence padding component 616 may apply padding values that may correspond to the baseline signal level observed during ambient sampling phases. In some implementations, the sequence padding component 616 may incorporate mechanisms to identify sequences with missing data points and may apply zero-padding to fill gaps in the temporal sequence.

[0133]

[0094] In some examples, the latent representation weighting component 618 may be configured as or otherwise support a means for weighting the task-relevant latent representation during training to emphasize minority class signals in imbalanced datasets while suppressing overrepresentation of majority class signals. In some implementations, the latent representation weighting component 618 may determine class weights dynamically based on the frequency distribution of task labels in the training dataset. In some implementations, the latent representation weighting component 618 may incorporate mechanisms to adjust the weighting scheme iteratively during training to account for changes in class representation across epochs. In some implementations, the latent representation weighting component 618 may apply scaling factors to the task-relevant latent representation that may amplify signals associated with underrepresented classes while attenuating signals linked to overrepresented classes.

[0134]

[0095] In some examples, the decoder training component 620 may be configured as or otherwise support a means for training the decoder to reconstruct the sensor data by combining the task-relevant latent representation and the confounder-related latent representation while preserving temporal causality in the reconstructed signals. In some 24 #18994037vlSTR-027W001

[0135] implementations, the decoder training component 620 may incorporate mechanisms to align reconstructed signals with the original temporal sequence by applying time-series alignment techniques. In some implementations, the decoder training component 620 may accommodate reconstruction processes that may handle varying levels of noise in the input data to maintain fidelity in the reconstructed signals. In some implementations, the decoder training component 620 may include transposed convolutional layers that may reconstruct highdimensional sensor data from the latent representations while retaining the temporal structure of the original signals.

[0136]

[0096] FIG. 7 shows a diagram of a system 700 including a device 702 that supports disentangling confounders from volatile organic compound data with adversarial learning in accordance with aspects of the present disclosure. The device 702 may be an example of or include the components of a database server or an apparatus 502 as described herein. The device 702 may include components for bi-directional data communications including components for transmitting and receiving communications, including a confounder disentanglement component 704, an I / O controller 706, a database controller 708, memory 710, a processor 712, and a database 714. These components may be in electronic communication via one or more buses (e g., bus 716).

[0137]

[0097] The confounder disentanglement component 704 may be an example of a confounder disentanglement component 506 or 602 as described herein. For example, the confounder disentanglement component 704 may perform any of the methods or processes described above with reference to FIGS. 5 and 6. In some cases, the confounder disentanglement component 704 may be implemented in hardware, software executed by a processor, firmware, or any combination thereof.

[0138]

[0098] The I / O controller 706 may manage input signals 718 and output signals 720 for the device 702. The I / O controller 706 may also manage peripherals not integrated into the device 702. In some cases, the I / O controller 706 may represent a physical connection or port to an external peripheral. In some cases, the I / O controller 706 may utilize an operating system such as iOS®, ANDROID®. MS-DOS®, MS-WINDOWS®, OS / 2®, UNIX®.

[0139] LINUX®, or another known operating system. In other cases, the I / O controller 706 may represent or interact with a modem, a keyboard, a mouse, a touchscreen, or a similar device. In some cases, the I / O controller 706 may be implemented as part of a processor. In some cases, a user may interact with the device 702 via the I / O controller 706 or via hardware components controlled by the I / O controller 706.

[0140] 25 #18994037vlSTR-027W001

[0141]

[0099] The database controller 708 may manage data storage and processing in a database 714. In some cases, a user may interact with the database controller 708. In other cases, the database controller 708 may operate automatically without user interaction. The database 714 may be an example of a single database, a distributed database, multiple distributed databases, a data store, a data lake, or an emergency backup database.

[0142]

[0100] Memory 710 may include random- access memory (RAM) and read-only memory (ROM). The memory 710 may store computer-readable, computer-executable software including instructions that, when executed, cause the processor to perform various functions described herein. In some cases, the memory 710 may contain, among other things, a basic input / output system (BIOS) which may control basic hardware or software operation such as the interaction with peripheral components or devices.

[0143]

[0101] The processor 712 may include an intelligent hardware device, (e.g., a general-purpose processor, a DSP, a central processing unit (CPU), a microcontroller, an ASIC, an FPGA, a programmable logic device, a discrete gate or transistor logic component, a discrete hardware component, or any combination thereof). In some cases, the processor 712 may be configured to operate a memory array using a memory controller. In other cases, a memory controller may be integrated into the processor 712. The processor 712 may be configured to execute computer-readable instructions stored in a memory7710 to perform various functions (e.g., functions or tasks supporting disentangling confounders from volatile organic compound data with adversarial learning).

[0144]

[0102] FIG. 8 shows a flowchart illustrating a method 800 that supports disentangling confounders from volatile organic compound data with adversarial learning in accordance with various aspects of the present disclosure. The operations of the method 800 may be implemented by one or more components of a networked computing system as described herein. For example, the operations of the method 800 may be performed by a confounder disentanglement component as described with reference to FIGS. 5 through 7. In some examples, one or more components of a networked computing system may execute a set of instructions to control the functional elements of the component(s) to perform the described functions. Additionally or alternatively, the one or more components of a networked computing system may perform aspects of the described functions using special-purpose hardware.

[0145]

[0103] At 802, the method 800 may include receiving sensor data including volatile organic compound signals and confounder-related signals. The operations of 802 may be performed in accordance with examples as disclosed herein. In some examples, aspects of the 26 #18994037vlSTR-027W001

[0146] operations of 802 may be performed by a sensor data receiving component 604 as described with reference to FIG. 6.

[0147]

[0104] At 804, the method 800 may include encoding the sensor data into a task-relevant latent representation and a confounder-related latent representation, the encoding performed by an encoder configured to separate the volatile organic compound signals from the confounder-related signals. The operations of 804 may be performed in accordance with examples as disclosed herein. In some examples, aspects of the operations of 804 may be performed by an encoder configuration component 606 as described with reference to FIG. 6.

[0148]

[0105] At 806, the method 800 may include adversarially training a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation. The operations of 806 may be performed in accordance with examples as disclosed herein. In some examples, aspects of the operations of 806 may be performed by a confounder predictor training component 608 as described with reference to FIG. 6.

[0149]

[0106] At 808, the method 800 may include reconstructing the sensor data from the taskrelevant latent representation and the confounder-related latent representation, the reconstruction performed by a decoder configured to preserve the volatile organic compound signals and isolate the confounder-related signals. The operations of 808 may be performed in accordance with examples as disclosed herein. In some examples, aspects of the operations of 808 may be performed by a sensor data reconstruction component 610 as described with reference to FIG. 6.

[0150]

[0107] It should be noted that the methods described herein describe possible implementations, and that the operations and the steps may be rearranged or otherwise modified and that other implementations are possible. Furthermore, aspects from two or more of the methods may be combined.

[0151]

[0108] Aspect 1: A method for disentangling confounders from volatile organic compound data with adversarial learning, comprising: receiving sensor data including volatile organic compound signals and confounder-related signals; encoding the sensor data into a task-relevant latent representation and a confounder-related latent representation, the encoding performed by an encoder configured to separate the volatile organic compound signals from the confounder-related signals; adversarially training a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation; and reconstructing the sensor data from the task-relevant latent representation 27 #18994037vlSTR-027W001

[0152] and the confounder-related latent representation, the reconstruction performed by a decoder configured to preserve the volatile organic compound signals and isolate the confounder-related signals.

[0153]

[0109] Aspect 2: The method of aspect 1, further comprising normalizing the sensor data by adjusting the volatile organic compound signals and the confounder-related signals based on a reference baseline derived from ambient sampling prior to encoding the sensor data.

[0154]

[0110] Aspect 3: The method of any of aspects 1 through 2. further comprising truncating the sensor data to exclude uninformative temporal regions by retaining a fixed window encompassing baseline, exposure, and recovery phases prior to encoding the sensor data.

[0155]

[0111] Aspect 4: The method of any of aspects 1 through 3. further comprising applying zero-padding to the sensor data sequences shorter than a predefined length to maintain uniform tensor dimensions for batch processing prior to encoding the sensor data.

[0156]

[0112] Aspect 5: The method of any of aspects 1 through 4, further comprising weighting the task-relevant latent representation during training to emphasize minority7class signals in imbalanced datasets while suppressing overrepresentation of majority class signals.

[0157]

[0113] Aspect 6: The method of any of aspects 1 through 5. further comprising training the decoder to reconstruct the sensor data by combining the task-relevant latent representation and the confounder-related latent representation while preserving temporal causality in the reconstructed signals.

[0158]

[0114] Aspect 7: The method of any of aspects 1 through 6. wherein the encoder is configured to apply temporal convolutional layers to the sensor data to extract sequential features while reducing noise from the confounder-related signals.

[0159]

[0115] Aspect 8: The method of any of aspects 1 through 7, wherein the confounder predictor is trained to identify confounder-related signals by minimizing a loss function specific to the confounder-related latent representation.

[0160]

[0116] Aspect 9: The method of any of aspects 1 through 8, wherein the task-relevant latent representation is configured to retain volatile organic compound signals by suppressing features correlated with the confounder-related latent representation.

[0161]

[0117] Aspect 10: The method of any of aspects 1 through 9, wherein the reconstruction performed by the decoder includes aligning the temporal structure of the reconstructed sensor data with the original sensor data.

[0162]

[0118] Aspect 11 : The method of any of aspects 1 through 10, wherein the adversarial training includes dynamically adjusting the learning rate of the confounder predictor to enhance separation between the task-relevant and confounder-related latent representations.

[0163] 28 #18994037vlSTR-027W001

[0164]

[0119] Aspect 12: The method of any of aspects 1 through 11, wherein the encoder is configured to generate the task-relevant latent representation and the confounder-related latent representation by applying distinct activation functions to separate task-relevant and confounder-related features.

[0165]

[0120] Aspect 13: A system for disentangling confounders from volatile organic compound data with adversarial learning, comprising a processor; memory coupled with the processor; and instructions stored in the memory and executable by the processor to cause the system to perform a method of any of aspects 1 through 12.

[0166]

[0121] Aspect 14: A system for disentangling confounders from volatile organic compound data with adversarial learning, comprising at least one means for performing a method of any of aspects 1 through 12.

[0167]

[0122] Aspect 15: A non-transitory computer-readable medium storing code for disentangling confounders from volatile organic compound data with adversarial learning, the code comprising instructions executable by a processor to perform a method of any of aspects 1 through 12.

[0168]

[0123] The description set forth herein, in connection with the appended drawings, describes example configurations and does not represent all the examples that may be implemented or that are within the scope of the claims. The term “exemplary” used herein means “serving as an example, instance, or illustration,” and not “preferred” or “advantageous over other examples.” The detailed description includes specific details for the purpose of providing an understanding of the described techniques. These techniques, however, may be practiced without these specific details. In some instances, well-known structures and devices are shown in block diagram form in order to avoid obscuring the concepts of the described examples.

[0169]

[0124] In the appended figures, similar components or features may have the same reference label. Further, various components of the same type may be distinguished by following the reference label by a dash and a second label that distinguishes among the similar components. If just the first reference label is used in the specification, the description is applicable to any one of the similar components having the same first reference label irrespective of the second reference label.

[0170]

[0125] Information and signals described herein may be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips that may be referenced throughout the above

[0171] 29 #18994037vlSTR-027W001

[0172] description may be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0173]

[0126] The various illustrative blocks and modules described in connection with the disclosure herein may be implemented or performed with a general-purpose processor, a DSP, an ASIC, an FPGA or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices (e.g., a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration).

[0174]

[0127] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Other examples and implementations are within the scope of the disclosure and appended claims. For example, due to the nature of software, functions described herein can be implemented using software executed by a processor, hardware, firmware, hardwiring, or combinations of any of these. Features implementing functions may also be physically located at various positions, including being distributed such that portions of functions are implemented at different physical locations. Also, as used herein, including in the claims, “or” as used in a list of items (for example, a list of items prefaced by a phrase such as “at least one of’ or “one or more of’) indicates an inclusive list such that, for example, a list of at least one of A, B, or C means A or B or C or AB or AC or BC or ABC (i. e.. A and B and C). Also, as used herein, the phrase “based on” shall not be construed as a reference to a closed set of conditions. For example, an exemplary step that is described as “based on condition A” may be based on both a condition A and a condition B without departing from the scope of the present disclosure. In other words, as used herein, the phrase “based on” shall be construed in the same manner as the phrase “based at least in part on.”

[0175]

[0128] Computer-readable media includes both non-transitory computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A non-transitory' storage medium may be any available medium that can be accessed by a general purpose or special purpose computer. By way of example, and not limitation, non-transitory computer-readable media can comprise RAM,

[0176] 30 #18994037vlSTR-027W001

[0177] ROM, electrically erasable programmable read only memory (EEPROM), compact disk (CD) ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to carry or store desired program code means in the form of instructions or data structures and that can be accessed by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor. Also, any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. Disk and disc, as used herein, include CD. laser disc, optical disc, digital versatile disc (DVD), floppy disk and Blu-ray disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above are also included within the scope of computer-readable media.

[0178]

[0129] The description herein is provided to enable a person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other variations without departing from the scope of the disclosure. Thus, the disclosure is not limited to the examples and designs described herein, but is to be accorded the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. STR-027W001CLAIMS1. A method for disentangling confounders from volatile organic compound data with adversarial learning, the method comprising:receiving sensor data including volatile organic compound signals and confounder-related signals;encoding the sensor data into a task-relevant latent representation and a confounder-related latent representation, the encoding performed by an encoder configured to separate the volatile organic compound signals from the confounder-related signals;adversarially training a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation; andreconstructing the sensor data from the task-relevant latent representation and the confounder-related latent representation, the reconstruction performed by a decoder configured to preserve the volatile organic compound signals and isolate the confounder-related signals.

2. The method of claim 1, further comprising normalizing the sensor data by adjusting the volatile organic compound signals and the confounder-related signals based on a reference baseline derived from ambient sampling prior to encoding the sensor data.

3. The method of claim 1, further comprising truncating the sensor data to exclude uninformative temporal regions by retaining a fixed window encompassing baseline, exposure, and recovery phases prior to encoding the sensor data.

4. The method of claim 1, further comprising applying zero-padding to sequence of the sensor data shorter than a predefined length to maintain uniform tensor dimensions for batch processing prior to encoding the sensor data.

5. The method of claim 1, further comprising weighting the task-relevant latent representation during training to emphasize minority class signals in imbalanced datasets while suppressing overrepresentation of majority class signals.32 #18994037vlSTR-027W0016. The method of claim 1, further comprising training the decoder to reconstruct the sensor data by combining the task-relevant latent representation and the confounder-related latent representation while preserving temporal causality in the reconstructed signals.

7. The method of claim 1, wherein the encoder is configured to apply temporal convolutional layers to the sensor data to extract sequential features while reducing noise from the confounder-related signals.

8. The method of claim 1, wherein the confounder predictor is trained to identify confounder-related signals by minimizing a loss function specific to the confounder-related latent representation.

9. The method of claim 1, wherein the task-relevant latent representation is configured to retain volatile organic compound signals by suppressing features correlated with the confounder-related latent representation.

10. The method of claim 1, wherein the reconstruction performed by the decoder includes aligning a temporal structure of the reconstructed sensor data with the received sensor data.

11. The method of claim 1, wherein the adversarial training includes dynamically adjusting a learning rate of the confounder predictor to enhance separation between the task-relevant and confounder-related latent representations.

12. The method of claim 1, wherein the encoder is configured to generate the task-relevant latent representation and the confounder-related latent representation by applying distinct activation functions to separate task-relevant and confounder-related features.

13. A system configured for disentangling confounders from volatile organic compound data with adversarial learning, comprising:a processor;memory coupled with the processor; andinstructions stored in the memory' and executable by the processor to cause the system to:33 #18994037vlSTR-027W001receive sensor data including volatile organic compound signals and confounder-related signals;encode the sensor data into a task-relevant latent representation and a confounder-related latent representation, the encoding performed by an encoder configured to separate the volatile organic compound signals from the confounder- related signals;adversanally train a confounder predictor to extract the confounder-related signals from the confounder-related latent representation while training the encoder to suppress the confounder-related signals in the task-relevant latent representation; and reconstruct the sensor data from the task-relevant latent representation and the confounder-related latent representation, the reconstruction performed by a decoder configured to preserve the volatile organic compound signals and isolate the confounder-related signals.

14. The system of claim 13, wherein the instructions are further executable by the processor to cause the system to: normalize the sensor data by adjusting the volatile organic compound signals and the confounder-related signals based on a reference baseline derived from ambient sampling prior to encoding the sensor data.

15. The system of claim 13, wherein the instructions are further executable by the processor to cause the system to: truncate the sensor data to exclude uninformative temporal regions by retaining a fixed window encompassing baseline, exposure, and recovery phases prior to encoding the sensor data.

16. The system of claim 13, wherein the instructions are further executable by the processor to cause the system to: apply zero-padding to sequences of the sensor data shorter than a predefined length to maintain uniform tensor dimensions for batch processing prior to encoding the sensor data.

17. The system of claim 13, wherein the instructions are further executable by the processor to cause the system to: weight the task-relevant latent representation during training to emphasize minority class signals in imbalanced datasets while suppressing overrepresentation of majority class signals.34 #18994037vlSTR-027W00118. The system of claim 13, wherein the instructions are further executable by the processor to cause the system to: train the decoder to reconstruct the sensor data by combining the taskrelevant latent representation and the confounder-related latent representation while preserving temporal causality in the reconstructed signals.

19. The system of claim 13, wherein the encoder is configured to apply temporal convolutional layers to the sensor data to extract sequential features while reducing noise from the confounder-related signals.35 #18994037vl

Images