Systems for analysing a gaseous biological sample using terahertz time-domain spectroscopy

Abstract

According to one aspect, the present description relates to a system (100) for analysing a gaseous biological sample, the system comprising a device (110) for collecting the gaseous biological sample; a time-domain spectroscopy measurement device (120) comprising a gas analysis cell (122) configured to receive the gaseous biological sample and electromagnetic detection means configured to detect at least one first sample time trace (Es(t)), each sample time trace resulting from coherent detection of a sample beam arising from the gas analysis cell traversed by a THz excitation beam; a processing unit (130) comprising a pre-trained machine learning module for detecting at least one state of a subject, wherein the processing unit is configured to calculate a sample estimator on the basis of the at least one sample time trace and to determine, from the at least one estimator and using the machine learning module, the at least one state of the subject.

Description

Description Title of the invention: Systems for analyzing a gaseous biological sample by terahertz spectroscopy in the time domain Technical field

[0001] The present description relates to systems for analyzing a gaseous biological sample by terahertz (THz) spectroscopy in the time domain, in particular for diagnostic purposes. The present description relates more particularly to systems for analyzing the exhaled air of a patient (human or animal). State of the art

[0002] The analysis of volatile organic compounds (VOCs) found in a patient's exhaled air (breath) has many applications, such as the diagnosis of pathologies such as asthma, diabetes or certain cancers (see [Ref. 1]).

[0003] Two main methodologies are known today for carrying out such an analysis.

[0004] A first methodology consists of systematically analyzing all volatile organic compounds, for example using mass spectroscopy (see [ref. 2]). However, an instrument for mass spectroscopy is expensive.

[0005] A second methodology consists of focusing the analysis on one or a few volatile compounds using an "electronic nose" which is much less expensive but which cannot be interested in all 3500 volatile organic compounds present in the breath (see [Ref. 3]).

[0006] Recently, it has been demonstrated (see [Ref. 4]) that terahertz (THz) spectroscopy in the time domain, and in particular thanks to a super-resolution method as described in [Ref. 4] can prove particularly effective in measuring the relative concentrations of hundreds of volatile components of a biological sample. A spectrometer as described in [Ref. 4] thus allows breath analysis with very good sensitivity.

[0007] However, the application of such a method to diagnosis is not envisaged. Indeed, applying the teaching of [Ref. 4] to diagnosis would require knowledge of the spectra of all the components specific to a pathology.

[0008] In [Ref. 5] it was also proposed, using terahertz (THz) spectroscopy in the time domain, to determine spectral lines or combinations of lines of acetone in exhaled air, for patients with diabetes and healthy patients, for diagnostic purposes.

[0009] To extend the teaching of [Ref.5] to the diagnoses of different pathologies, it would therefore be necessary to have a priori knowledge of the spectra of the components volatile organic compounds as well as knowledge of the VOC bouquets that are markers of a given pathology. However, this a priori knowledge would be laborious to establish for a large number of pathologies.

[0010] The present description proposes a system for analyzing a gaseous biological sample and in particular a breath analysis system, which allows rapid diagnosis of predefined pathologies without a priori knowledge of either the bouquets of VOCs markers of a given pathology, or the spectra of volatile organic components. Summary

[0011] In this description, the term "comprise" means the same as "include", "contain", and is inclusive or open and does not exclude other elements not described or shown. Furthermore, in this description, the term "approximately" or "substantially" means the same as "having a margin less than and / or more than 10%, for example 5%", of the respective value.

[0012] The present description relates, according to a first aspect, to a system for analyzing a gaseous biological sample comprising: - a device for collecting said gaseous biological sample; - a time domain spectroscopy measuring device (120) comprising: - a gas analysis cell configured to receive the collected gaseous biological sample; - electromagnetic emission means configured to emit into the gas analysis cell, a substantially collimated THz excitation beam; - electromagnetic detection means configured to detect at least a first sample time trace, each sample time trace resulting from a coherent detection of a THz sample beam originating from the gas analysis cell crossed by the THz excitation beam; - a processing unit comprising a pre-trained machine learning module for detecting at least one state of a subject, the processing unit being configured to: - calculating a sample estimator from said at least one sample time trace; - determining from said estimator and by means of the machine learning module, said at least one state of the subject.

[0013] The THz excitation beam comprises in a known manner (see for example [Ref. 4]) electromagnetic pulses emitted with a given period and a spectrum formed by a comb of predetermined frequencies. The THz sample beam results from the convolution of the THz excitation beam with a function which depends on pa- physical parameters characteristic of the gaseous biological sample, for example the absorption coefficient and / or the refractive index of the sample, this function being called the transfer function of the sample.

[0014] Coherent THz sample beam detection is a detection sensitive to the effect of the gaseous biological sample on the amplitude and phase of the incident THz excitation beam.

[0015] As is known, such coherent detection comprises, in exemplary embodiments, the generation of an ultrashort (wide bandwidth) terahertz pulse from an even shorter femtosecond optical pulse, emitted for example by a Ti-sapphire laser. The optical pulse is first split to provide an optical probe pulse whose path length is adjusted using an optical delay line. The optical probe pulse illuminates a teraherz detector which is sensitive to the electric field of the resulting THz sample beam at the time the optical probe pulse is delivered to the detector. By varying the path length traveled by the optical probe pulse, a time trace is thus measured as a function of time. The response of a sample can be calibrated by means of a reference time trace, obtained under the same experimental conditions but with the sample removed, for example.

[0016] The applicants have shown that the system for analyzing a gaseous biological sample according to the first aspect allows rapid diagnosis of predefined pathologies without a priori knowledge of either the bouquets of VOCs markers of a given pathology, or the spectra of the volatile organic components thanks to processing of the sample time traces and not the spectra. This processing is done by means of a pre-trained machine learning module, also called a "prediction module" in the present description. The use of such a prediction module also makes it possible, in exemplary embodiments, to increase the prediction performance of the module by continuing to train the prediction module with new data acquired during clinical use of the analysis system.

[0017] According to one or more exemplary embodiments, the time domain spectroscopic measurement device is configured to further detect at least one first reference time trace; and said sample estimator is calculated from said at least one sample time trace and said at least one reference time trace.

[0018] According to one or more exemplary embodiments, the detection means are configured to detect a plurality of sample time traces, the sample estimator being calculated from said plurality of sample time traces.

[0019] In exemplary embodiments, a sample estimator is for example obtained from an average of said sample time traces.

[0020] According to one or more exemplary embodiments, the machine learning module is pre-trained using a convolutional neural network.

[0021] According to one or more exemplary embodiments, the collection device is configured for collecting breath from a patient and includes a breath collection tube.

[0022] In exemplary embodiments, the collection device further comprises carbon dioxide detection means configured to detect carbon dioxide in the collected breath, a first solenoid valve and means for controlling the first solenoid valve configured to open the solenoid valve when the detected carbon dioxide exceeds a predetermined threshold value, the collected breath then being able to be sucked towards the gas analysis cell.

[0023] In exemplary embodiments, the collection device further comprises an intermediate enclosure configured to receive at least a fraction of the collected breath, a second solenoid valve and means for controlling the second solenoid valve configured to open said second solenoid valve, the collected breath then being able to be sucked towards the gas analysis cell and close said second solenoid valve when a pressure in the gas analysis cell reaches a predetermined threshold value.

[0024] According to one or more exemplary embodiments, the collection device is configured for collecting a liquid biological sample and comprises means for producing, from the liquid biological sample, a gaseous biological sample. A liquid biological sample is, for example, a sample of urine, perspiration or saliva.

[0025] According to one or more exemplary embodiments, the electromagnetic emission means and the electromagnetic detection means each comprise a teraherz beacon, each of the teraherz beacons comprising an antenna, a reflecting mirror, for example a parabolic mirror and a support configured to securely hold said antenna and said reflecting mirror.

[0026] It is thus possible to obtain a collimated teraherz beam from a diverging point source, without using a teraherz lens which generates signal losses. The orientation of the teraherz beam is also facilitated. In particular, it is possible to provide for the antenna and / or for the deflecting mirror a plate configured for the positioning and / or orientation of said antenna or said deflecting mirror, which presents another advantage compared to the use of a teraherz lens.

[0027] According to one or more exemplary embodiments, the system for analyzing a gaseous biological sample further comprises: - a database comprising, for a set of subjects, at least one sample time trace associated with a state of the subject, and - a training module for training the learning module au- automated from database data; - the processing unit being further configured to send said at least one sample time trace generated by the time domain spectroscopy measuring device to said database.

[0028] The database is advantageously stored in a dematerialized form (“cloud” according to the Anglo-Saxon expression).

[0029] Such a system further allows for an update of the machine learning module during clinical use of the analysis system. Brief description of the figures

[0030] Other advantages and characteristics of the technique presented above will appear on reading the detailed description below, made with reference to the figures in which: - [Fig.lA], an example of a system for analyzing a gaseous biological sample according to the present description; - [Fig.lB], diagrams illustrating an example of a temporal shape of an excitation beam and an example of a spectrum of an excitation signal, in an analysis system according to the present description; - [Fig.lC], a diagram illustrating an example of a time trace measured by means of an analysis system according to the present description; - [Fig.lD], diagrams showing examples of sample and reference time traces; - [Fig.2A], a diagram illustrating in more detail a gas analysis cell in a measuring device of an analysis system according to the present description, with a first example of a collection device; - [Fig.2B], a diagram illustrating a second example of a collection device in an analysis system according to the present description; - [Fig.3], diagrams illustrating an example of a terahertz beacon in an analysis system according to the present description; - [Fig.4], a figure showing in a three-quarter view, a system for analyzing a gaseous biological sample according to the present description, mounted in a rack; - [Fig.5], a functional diagram illustrating different modules of a system for analyzing a gaseous biological sample according to the present description, in an exemplary implementation according to the present description; - [Fig.6], a functional diagram illustrating different modules of an analysis system configured for training the learning module.

[0031] In the various embodiments which will be described with reference to the figures, similar or identical elements bear the same references. Detailed description

[0032] In the following detailed description, only certain embodiments are described in detail to ensure clarity of the disclosure, but these examples are not intended to limit the general scope of the principles emerging from this description.

[0033] The various embodiments and aspects described in the present description may be combined or simplified in multiple ways. In particular, the steps of the various methods may be repeated, interchanged, executed in parallel, unless otherwise specified.

[0034] [Fig.1A] illustrates a first example of a system 100 for analyzing a gaseous biological sample according to the present description.

[0035] The analysis system comprises a sample collection device 110 and a time domain spectroscopy (or TDS) measuring device 120. The time domain spectroscopy device 120 comprises a gas analysis cell 122 configured to receive the collected gaseous biological sample, an example of which will be described in more detail using [Fig. 2A]. The time domain spectroscopy device 120 further comprises electromagnetic emission means 124, 126 configured to emit, into the gas analysis cell, a substantially collimated THz excitation beam and detection means 128 configured to detect at least a first sample time trace resulting from a coherent detection of a sample beam originating from the gas cell illuminated by the excitation beam.The sample beam results in practice from the convolution of the excitation beam with a transfer function of the gaseous sample to be analyzed. The analysis system 100 also comprises a processing unit 130 comprising a pre-trained machine learning module for detecting at least one state of a subject. As will be explained in more detail later, the processing unit 130 is configured to calculate an estimator from said at least one sample time trace and determine from said estimator and by means of the machine learning module, said at least one state of the subject. In the example of [Fig. 1A], the analysis system 100 further comprises a vacuum pump 150 configured to empty the gas analysis cell 122 via a tube and solenoid valves, as will be described in more detail by means of [Fig.2A] and a set 140 of microcontroller(s) and / or sensor(s), in connection with the vacuum pump 150, the gas analysis cell 122 and the processing unit 130, as will be described in more detail later.

[0036] In operation, as illustrated in [Fig.lA] and according to embodiment examples, the gaseous sample passes through the collection device 110, for example a device for collecting a patient's breath; a portion of the sample passes through the collection device and a portion is aspirated via a bypass capillary 111. The path of the gaseous biological sample is indicated by simple arrows in [Fig.lA]. The breath collection comprises, for example, a tube, for example a PTFE tube. The tube has, for example, a diameter of between approximately 10 mm and approximately 15 mm and a length of between approximately 20 cm and approximately 40 cm. The tube in the example of the breath collection device 110 illustrated in [Fig.lA] comprises an end open to the open air to allow the circulation of the breath. This makes it possible, on the one hand, to avoid causing a suction effect on the subject, the cell 122 being under vacuum, and on the other hand to allow the collection in the cell of the portion of the breath that is to be analyzed.For example, if we want to analyze the air coming from the bottom of the lungs, the aspiration of the breath into the cell can be triggered towards the end of the exhalation. On the other hand, if we want to analyze the air coming from the mouth or the bronchi, we can trigger the aspiration as early as possible in the exhalation.

[0037] The capillary 111 opens according to examples into an intermediate enclosure (not shown in [Fig. 1A] but described in more detail in [Fig. 2A]) then into the gas analysis cell 122 of the time domain spectroscopy measuring device 120, where it is analyzed.

[0038] As explained previously, for the analysis of the gaseous sample, the time domain spectroscopy device 120 comprises one of the electromagnetic emission / detection means 124, 125, a THz transmitter 126 as well as a THz receiver 128.

[0039] THz transmitters / receivers configured for transmitting / receiving pulses in the THz frequency band (i.e., between about 0.2 Thz and about 8 Thz) are known to those skilled in the art.

[0040] In exemplary embodiments, the transmitter 126 and receiver 128, referred to herein as "THz headlights," comprise an antenna and a collimator, such as a lens or parabolic mirror. Examples of THz headlights will be described in more detail using [Fig. 3].

[0041] The electromagnetic emission means 124 comprise, for example and in a known manner, a pulsed femtosecond laser, for example a Ti-sapphire laser, and a delay line; in operation, a femtosecond pulse is separated into a first pulse directed towards the antenna of the THz transmitter 126 and into a second pulse, or optical probe pulse, directed towards the antenna of the THz receiver 128 after having undergone a variable time delay thanks to the delay line.

[0042] The femtosecond laser is for example a pulsed frequency comb laser configured to excite the antenna of the THz 126 lighthouse (transmitter), the femto- pulses seconds being conveyed to the antenna, for example, by an optical fiber 127. The antenna of the Thz lighthouse 126 comprises, for example, a semiconductor element and polarized electrodes. Illuminating the semiconductor element at an energy greater than the bandgap energy makes it possible to generate free carriers. The semiconductor therefore changes from an insulating state to a conducting state, generating an electric current between the polarized electrodes of the antenna. This results in the emission of a THz pulse. The THz pulse is directed by means of the collimator in the cell 122 through the gaseous biological sample to be studied to form a THz excitation beam.

[0043] An example of a THz excitation beam is shown in [Fig. 1B] (left diagram). It comprises, for example, electromagnetic pulses 10 emitted periodically with a period T; it is characterized by a spectrum 11 (right diagram) consisting of a frequency comb. The pulses 10 have a duration between several hundred femtoseconds and a few picoseconds. The repetition frequency (equal to 1 / T) of the laser can vary between 1 GHz to a frequency lower than 1 Hz.

[0044] Each THz pulse after passing through the sample is directed towards the antenna of the THz 128 lighthouse (receiver) by means of the THz 128 lighthouse collimator to form a THz sample beam which results from the convolution of the THz excitation beam with a transfer function associated with the gaseous biological sample and which depends on characteristic parameters of the sample, in particular its absorption.

[0045] The electric field of the terahertz pulses is measured at the antenna of the THz lighthouse 128 illuminated simultaneously by the optical probe pulse routed to the antenna of the THz lighthouse 128 for example by means of an optical fiber 129, said optical probe pulse having undergone a delay relative to the optical pulse sent to the antenna of the THz lighthouse 126 (transmitter) thanks to the delay line of the electromagnetic emission means 124. The electrical signal generated at the antenna can be amplified, then detected by electrical detection means 125. The electrical detection means 125 thus measure an electric field as a function of time, on scales ranging from the femtosecond to several hundred picoseconds or even nanoseconds. Coherent detection of the THz sample beam is thus obtained.

[0046] Thus, as illustrated in [Fig.lC], the time-dependent measurement can be achieved by means of photoconductive or electro-optical sampling, as explained for example in [Ref. 4]. The delay line comprises for example mirrors mounted on a motorized translation stage, introducing a delay having a maximum time excursion t max . Time sampling can also be achieved by beating two frequency combs whose repetition rate is le- different management (so-called HASSOPS technique). The time interval between two measurements is called the sampling period t s , with sampling frequency f s = l / t s . The time excursion t max is the time range over which the electric field measurement is carried out; this measurement of the electric field as a function of time is called the sample time trace Es(t). A typical sampling period t sis between about 10 and 50 fs. It depends on the rate at which the femtosecond laser fires optical pulses. The ability to directly measure the electric field of the THz pulse rather than the averaged energy gives access to both the phase and amplitude of the waveform, and thus provides information on the absorption coefficient and refractive index of the sample. In this case, the measurement of the electric field as a function of time under reference conditions, for example under the same experimental conditions, but without the sample, is called a reference time trace Eref(t).

[0047] [Fig.lD] illustrates examples of a sample time trace Es(t) and a reference time trace Eref(t) thus detected.

[0048] The processing unit 130 receives the time traces generated by the electrical detection means 125.

[0049] The processing unit 130 may comprise one or more computers or calculation units. More generally, when in the present description, reference is made to calculation or processing steps for the implementation in particular of method steps, it is understood that each calculation or processing step may be implemented by software, hardware, firmware, microcode or any appropriate combination of these technologies. When software is used, each calculation or processing step may be implemented by computer program instructions or software code. These instructions may be stored or transmitted to a storage medium readable by a computer (or calculation unit) and / or be executed by a computer (or calculation unit) in order to implement these calculation or processing steps.

[0050] The processing unit communicates with the TDS 120 to send requests and retrieve the time traces in order to record and process them. The processing unit can also interact with the assembly 140 comprising microcontroller(s) and / or sensor(s). For example, a microcontroller can automate a valve circuit (See [Fig.2A]) and / or have an action for controlling the vacuum pump 150. One or more sensors can measure different parameters (pressure in the cell, temperature, CO2 in the patient's breath, etc.). The processing unit can then send requests to the sensors or a setpoint to the microcontroller and the values ​​measured by the sensors can be sent to the processing unit.

[0051] [Fig.2A] illustrates in more detail and according to an example, the operation of a system for analyzing a biological sample at the gas analysis cell 122.

[0052] In this example, the collection device 110 comprises a breath collection tube 112 and a carbon dioxide (CO2) sensor 115 arranged for example at the inlet of the tube, configured to measure the CO2 in the breath of a patient. From a predetermined threshold value of CO2, the processing unit can control the opening of a solenoid valve 212 arranged on the bypass conduit 111 to suck the breath towards a tubular enclosure 220 of the gas analysis cell 122. The tubular enclosure 220 is for example a stainless steel tube with connectors to connect the different gas inlets and outlets and the sensors.

[0053] A solenoid valve 214 may be provided to open until a pressure setpoint in the enclosure 220 is met.

[0054] An intermediate enclosure 215 may also be provided to collect the breath during sampling. This makes it possible, for example, to collect a fraction of the breath, for example the entire alveolar fraction of the breath, i.e. the air present in the lungs. A valve 212 makes it possible to collect the desired fraction of the breath in the intermediate enclosure 215.

[0055] Opening the valve 214 then allows the tubular enclosure 220 to be filled up to a pressure chosen for the analysis; for example, if a measurement is to be made at 10 mbar, the tubular enclosure is only filled with 10 mbar of gas included in the intermediate enclosure 215.

[0056] When emptying of the enclosure 220 is requested, a solenoid valve 222 can open to empty the enclosure 220.

[0057] A heating element 230, for example a heating wire, may be provided to regulate the temperature of the cell. A pressure gauge may be provided to measure the pressure in the cell.

[0058] In the example of [Fig.2A], the THz headlights (transmitter 255, receiver 265) are placed in enclosures 250, 260, opposite windows 253, 263 respectively, for example Brewster angle inclined windows. The windows 253, 263 are carried by supports which can be mechanically connected to the enclosure 220 by means of parts 254, 264, for example cable ties.

[0059] The enclosures 250 and 260, for example plexiglass boxes, can be configured to allow the atmosphere between the antennas of the THz headlights 255, 265 and the windows 253, 263 to be purged in order to get rid of the water present in the air and which could disturb the measurement. Indeed, between the antenna of the THz headlight and the window, the beam propagates in the open air, which is naturally charged with water. Since water is visible in the THz band, it introduces a bias in the measurement signal. Thus, filling the enclosures 250 and 260 with an inert gas, for example invisible nitrogen in the THz band, allows the air and water it contains to be expelled to saturate the atmosphere of the enclosures with nitrogen and generate an inert atmosphere).

[0060] [Fig.2B] illustrates another example of a collection device 210 configured to receive a liquid sample, for example a urine, saliva or perspiration sample.

[0061] Such a collection device may be connected to a tubular enclosure 220 of a gas analysis cell 122 as shown in [Fig.2A], replacing the elements 110, 111, 112, 115, 212, 214, 215.

[0062] The collection device 210 comprises a test tube 211 provided with a vacuum connection (not shown), a needle valve 213 for regulating the flow entering the gas analysis cell, and a manual valve 217 for completely closing the passage to the cell.

[0063] In operation, a sample purification step can first be carried out. For this, the sample can be frozen in the test tube using, for example, a liquid nitrogen bath, then a vacuum is created in the test tube, thanks to the tubular enclosure 220 which is under vacuum. This removes the air present in the test tube.

[0064] Once this step is completed, valves 213, 217 can be closed.

[0065] To take the sample, it is possible, for example, to open the manual valve 217 and to control the opening of the needle valve. Thus, the liquid sample present in the test tube 211 is subjected to vacuum, vaporizes and is sucked into the tubular enclosure 220.

[0066] As the tubular enclosure 220 fills, it is possible to monitor the evolution of the pressure in the tubular enclosure 220, and to close the valves 213, 217 once the quantity (measured in pressure) of gaseous biological sample to be measured is reached.

[0067] [Fig. 3] shows diagrams illustrating in different views an example of a terahertz beacon in an analysis system according to the present description.

[0068] The THz headlight illustrated in [Fig. 3] comprises an antenna 310 and a plate 320 for adjusting the position and / or orientation of the antenna 310, a collector 350, for example a parabolic mirror, and a plate 360 ​​for adjusting the position and / or orientation of the parabolic mirror 350, a support part 370 for mechanically connecting the antenna and the parabolic mirror to each other. In the example of [Fig. 3], the assembly formed by the antenna 310 and the plate 320 is fixed to the support part 370 by means of a mechanical interface 330. Furthermore, a foot 380 makes it possible to fix the THz headlight in the enclosure (250, 260, [Fig. 2A]) in which it is to be fixed.

[0069] [Fig.4] represents a figure showing in a three-quarter view, a system for analyzing a gaseous biological sample according to the present description, mounted in a bay 400. In particular, one can observe in [Fig.4] the gas analysis cell 122, the device for collecting the gaseous biological sample, for example a breath collection device 110, the electromagnetic emission means 124 and the electrical detection means 125, the processing unit 130, the vacuum pump 150.

[0070] [Fig.5] is a diagram illustrating in more detail modules of a system for analyzing a gaseous biological sample according to the present description, in examples of implementation.

[0071] The gaseous biological sample from the patient is collected by a collection device, for example a breath collection device 110 as described by means of [Fig. 2A], or a collection device 210 of a liquid sample as described by means of [Fig. 2B].

[0072] The gaseous biological sample is sent to the time domain spectroscopic measuring device 120 to perform the THz spectroscopic measurement, for example a device as described by means of [Fig.lA] and [Fig.2A]. The measured information is a time trace, i.e. the electric field measured as a function of time.

[0073] In operation, a reference time trace can be associated with each measured sample time trace. The reference time trace is measured when there is no sample in the cell. The reference time trace can be measured periodically (e.g., once a week, daily, or between patients).

[0074] For the same sample / reference, a set of time traces can be collected in order to increase the signal-to-noise ratio.

[0075] The time traces can be corrected by means of a correction module 510 in order to further increase the signal-to-noise ratio. An example of correction is a rectification of the offset of the traces relative to each other induced by the measurement system. An average can be taken over all the time traces, after correction. For example, between approximately 500 time traces and approximately 1500 time traces, for example approximately 1000 time traces, are recovered per sample, and per reference.

[0076] In operation, the patient and reference time traces can then be processed by a preprocessing module 520 in order to put them in the format suitable for interpretation by a prediction module 530, also called a machine learning module in the present description. A sample estimator is then obtained from which the prediction module 530 can predict the patient's membership class (for example, sick patient or healthy patient, or assessment of the risk of developing a pathology as a percentage, etc.). The membership class is then processed by a healthcare professional.

[0077] A system for analyzing a gaseous biological sample as illustrated in [Fig.4] can thus be made available to a user for clinical use.

[0078] [Fig.6] illustrates a system 600 configured for pre-training the prediction module 530 of the processing unit of an analysis system according to the present description.

[0079] The system may comprise all of the elements of an analysis system according to the present description; only the processing unit 630 is configured specifically to train the prediction module 530.

[0080] Thus, the system 600 comprises a collection device (not illustrated in [Fig. 6]) for generating a gaseous biological sample for a given patient and a time domain spectroscopy measuring device 120 for performing the measurement by THz spectroscopy, for example a device as described by means of [Fig. 1A] and [Fig. 2A]. A plurality of time traces are generated, as described previously, and can be corrected by means of the correction module 510.

[0081] In parallel with these measurements, the patient takes a test depending on the pathology studied in order to have information on his class (for example sick patient or healthy patient, type 1 or 2 diabetes; in the case of long-term follow-up on a cohort, development of a targeted pathology, etc.). Thus for each sample we have an associated known class.

[0082] The measurements taken will allow the creation of the 540 database to begin. This will advantageously be a database recorded in a dematerialized manner, i.e. in the Cloud. The database includes, for example and without limitation, for each patient, a unique identification, specific data (age, gender, etc.), the time traces of the sample measured for the patient, the reference time traces, the class of belonging (sick patient or healthy patient, type 1 or 2 diabetes, time of the pathology developed, etc.).

[0083] Once the database has been created, the data can be sent to a preprocessing module 650. Its role is to prepare the data which will be entered into a training and evaluation module 660. Data preparation allows the information to be concentrated in order to improve predictions [Ref. 6].

[0084] Training consists of making the model learn on the data, that is to say finding the parameters allowing the model to have good prediction performance criteria on the data. Several models (multi-layer perceptron, recurrent neural network, ...) [Ref. 7][ Ref. 8] can thus be trained and then evaluated in order to retain the one which has the best performance. A model is a mathematical function with parameters which takes data as input and returns an output. Several types of models exist in the literature [Ref. 8][ Ref. 9]. An example of a model which can be tested is a convolutional neural network, for example network neural network capable of learning dependencies over long sequences or time series or “Long Short Term memory Neural Network” according to the Anglo-Saxon expression [Ref. 10].

[0085] After training, comes the evaluation phase; the different models can be compared using metrics (precision, false positive rate, false negative rate, etc.) [Ref. 11], depending on the risk of the pathology studied. An example of a metric is the recall which measures the rate of positive cases correctly predicted among the positive cases for the pathology studied. Indeed, during clinical use, false positive results can be eliminated in the following steps after the control test.

[0086] Finally, the best trained model will be put into production and accessible via a final interface between the prototype and the user to constitute the prediction module 530 ([Fig.5]). Via this interface, the user will be able to analyze the patient's breath and have its class membership.

[0087] In exemplary embodiments, if, during clinical use, the patient's class was previously known, their data could be fed into the database, as illustrated with reference to [Fig.5]. The new data will then be used to improve the model and update the prediction module.

[0088] Thus, in exemplary embodiments, the system for analyzing a gaseous biological sample further comprises a module 580 for updating the prediction module 530.

[0089] In these exemplary embodiments, the measurements made during the clinical study can make it possible to enrich the database 540 previously created to train the prediction model 530, for example a database recorded in a dematerialized manner.

[0090] The database 540 is accessible by the processing unit 130 of the analysis system to send the new data collected there. As described previously, the database 540 may contain patient information: a unique identification per patient, specific data (age, gender, etc.), the time traces Eref(t), Es(t), associated with the patient, the class of membership (sick or not sick; type 1 or 2 diabetes; etc.). The class of membership depends on the pathology studied. Eref(t) and Es(t) can be obtained for example from the average of the time traces as explained previously.

[0091] The data from the database 540 are sent to a preprocessing module 550. As explained with reference to [Fig.6], its role is to prepare the data which will be used during the training phase by a training and evaluation module 560 to improve the predictions.

[0092] Training, as described with reference to [Fig.6], consists of making the model learn on the data, that is to say finding the parameters allowing the model to have good performance criteria on the data. An example of a model that can be tested is the convolutional neural network. The evaluation phase, as explained previously, allows for the comparison of different models based on the risk of the pathology studied.

[0093] At the output of the training and evaluation module, information on the patient's class is generated. It is then possible to update the prediction module 530 of the analysis system.

[0094] Although described through a number of detailed exemplary embodiments, the analysis systems include various variations, modifications and improvements which will be apparent to those skilled in the art, it being understood that these various variations, modifications and improvements are within the scope of the invention, as defined by the following claims. References

[0095] Ref 1: AW Boots, et al. “Exhaled molecular fingerprinting in diagnosis and monitoring: validating volatile promises”. Trends in Molecular Medicine, 21(10):633-644, 2015.

[0096] Ref 2: A. Ulanowska et al. “The application of statistical methods using VOCs to identify patients with lung cancer”, Journal of Breath Research, 5(4):046008, 2011.

[0097] Réf.3 : M. Phillips et al. « Variation in volatile organic compounds in the breath of normal humans”, Journal of Chromatography B: Biomedical Sciences and Applications, 729(l-2):75-88, 1999.

[0098] Réf. 4 : EP 3 865 857

[0099] Réf. 5 : Y. Kistenev et al. « Diagnosis of Diabetes Based on Analysis of Exhaled Air by Terahertz Spectroscopy and Machine Laming”, Optics and Spectroscopy, 128(6), pp. 809 - 814, 2020.

[0100] Réf. 6 : Benhar, H., Idri, A., & Fernândez-Alemân, J. L. (2020). Data preprocessing for heart disease classification : A systematic literature review. Computer Methods and Programs in Biomedicine, 195, 105635.

[0101] Réf. 7 Kotsiantis, S. B., Zaharakis, I. D., & Pintelas, P. E. (2006). Machine learning: a review of classification and combining techniques. Artificial Intelligence Review, 26(3), 159-190.

[0102] Réf. 8 Craik, A., He, Y., & Contreras-Vidal, J. L. (2019). Deep learning for electroencephalogram (EEG) classification tasks: a review. Journal of neural engineering, 16(3), 031001.

[0103] Réf. 9 Kiranyaz, S., Ince, T., Abdeljaber, O., Avci, O., & Gabbouj, M. (2019, May). 1-D convolutional neural networks for signal processing applications. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 8360-8364). IEEE.

[0104] Réf. 10 Murat, F., Yildirim, O., Talo, M., Baloglu, U. B., Demir, Y., & Acharya, U. R. (2020). Application of deep learning techniques for heartbeats detection using ECG signals-analysis and review. Computers in biology and medicine, 120, 103726.

[0105] Réf. 11 Hossin, M., & Sulaiman, M. N. (2015). A review on evaluation metrics for data classification evaluations. International journal of data mining & knowledge management process, 5(2), 1.

Claims

Claims

1. System (100) for analyzing a gaseous biological sample comprising: - a device (110, 210) for collecting said gaseous biological sample; - a time domain spectroscopy measuring device (120) comprising: - a gas analysis cell (122) configured to receive the collected gaseous biological sample; - electromagnetic emission means (124, 126) configured to emit into the gas analysis cell, a substantially collimated THz excitation beam; - electromagnetic detection means (128, 125) configured to detect at least a first sample time trace (Es(t)), each sample time trace resulting from a coherent detection of a THz sample beam originating from the gas analysis cell crossed by the THz excitation beam; - a processing unit (130) comprising a pre-trained machine learning module (530) for detecting at least one state of a subject, the processing unit being configured to: - calculating a sample estimator from said at least one sample time trace; - determining from said estimator and by means of the machine learning module, said at least one state of the subject.

2. A system for analyzing a gaseous biological sample according to claim 1, wherein - the time domain spectroscopy measuring device (120) is configured to further detect at least one first reference time trace; and - said sample estimator is calculated from said at least one sample time trace and said at least one reference time trace.

3. A system for analyzing a gaseous biological sample according to any one of the preceding claims, wherein the detection means is configured to detect a plurality of sample time traces, the sample estimator being calculated from said plurality of sample time traces.

4. System for analyzing a gaseous biological sample according to one of any of the preceding claims, wherein the machine learning module (530) is pre-trained using a convolutional neural network.

5. A system for analyzing a gaseous biological sample according to any one of the preceding claims, wherein the collection device (110) is configured for collecting the breath of a patient and comprises a breath collection tube (112).

6. A system for analyzing a gaseous biological sample according to claim 5, wherein the collection device (110) further comprises carbon dioxide detection means (115) configured to detect carbon dioxide in the collected breath, a first solenoid valve (212) and means for controlling the first solenoid valve configured to open the solenoid valve when the detected carbon dioxide exceeds a predetermined threshold value, the collected breath then being able to be drawn towards the gas analysis cell (122).

7. System for analyzing a gaseous biological sample according to any one of claims 5 or 6, wherein the collection device (110) further comprises an intermediate enclosure (215) configured to receive at least a fraction of the collected breath, a second solenoid valve (214) and means for controlling the second solenoid valve configured to open said second solenoid valve, the collected breath then being able to be sucked towards the gas analysis cell (122) and close said second solenoid valve when a pressure in the gas analysis cell (122) reaches a predetermined threshold value.

8. System for analyzing a gaseous biological sample according to any one of claims 1 to 4, wherein the collection device (210) is configured for the collection of a liquid biological sample and comprises means for producing, from the liquid biological sample, a gaseous biological sample.

9. System for analyzing a gaseous biological sample according to any one of the preceding claims, in which the electromagnetic emission means and the electromagnetic detection means each comprise a THz headlight (255, 265), each of the THz headlights comprising an antenna (310), a deflecting mirror (350), for example a parabolic mirror and a support (370) configured to securely hold said antenna and said parabolic mirror.

10. System for analyzing a gaseous biological sample according to one of any of the preceding claims, further comprising: - a database (540) comprising, for a set of subjects, at least one sample time trace associated with a state of the subject, and - a training module (550) for training the machine learning module (530) using data from the database; - the processing unit (130) being further configured to send said at least one sample time trace generated by the time domain spectroscopy measuring device (120) to said database.

Images