Method and device for automatic disease state diagnosis

Abstract

A method of automatically diagnosing pneumonia in a patient includes obtaining values of two or more diagnostic parameters of the patient from a caregiver of the patient using an input / output interface device. The method includes applying, using a processor connected to an input / output interface, two or more diagnostic parameters to an electronic memory storing a plurality of pre-compiled pneumonia diagnostic models to identify an optimal diagnostic model for performing the diagnosis. The values of the two or more diagnostic markers are applied to the identified optimal diagnostic model to generate a diagnostic output. An input / output interface device is operated according to the diagnostic output to indicate to the caregiver the presence or absence of pneumonia in the patient. The caregiver may use the diagnostics to provide appropriate care for the patient. A pneumonia diagnostic model is from investigations of pneumonia-positive and non-pneumonia population of subjects.

Description

Technical Field

[0001] This invention relates to methods and apparatus for assisting medical personnel in diagnosing and managing patients with respiratory dysfunction, such as pneumonia.

[0002] Related applications This application claims priority to Australian Provisional Patent Application Nos. 2016903894 and 2016903896, filed on September 26, 2016, the disclosure of which is incorporated herein by reference in its entirety. Background Technology

[0003] Any reference to prior art methods, apparatus or documents shall not be regarded as evidence or acknowledgment that they have become common general knowledge or are part of common general knowledge.

[0004] Pneumonia is one of the leading causes of death among children under five worldwide. An estimated 905,059 children under five died from pneumonia globally in 2013 [Reference 1]. This accounts for 14% of the 6.3 million child deaths worldwide that year [1,2]. The United Nations (UN) recognized this problem and, through the Millennium Development Goal (MDG) 4 program, collaborated with countries worldwide to reduce under-five mortality by two-thirds between 1990 and 2015 [3-5].

[0005] Currently, diagnostic tools used to diagnose pneumonia include a stethoscope and auscultation, chest X-ray, chest CT imaging, blood analysis, pulse oximeter, and microbiology laboratory tests. There is no gold standard for diagnosing pneumonia. Even in hospitals, a definitive diagnosis of pneumonia in children, especially in its early stages, is very difficult. Lung aspiration biopsy may be the most effective method, but it is clearly impractical for clinical use. Clinical examination with a stethoscope and chest auscultation are the first steps in diagnosing pneumonia in children. Auscultation requires a skilled physician, and even then, it cannot provide a diagnosis with sufficient sensitivity or specificity. Chest X-rays are often used as an important reference standard to confirm a clinical diagnosis. However, X-rays may be insensitive to early pneumonia or when the site of lung lesions is not clearly visible on the image. Furthermore, in the presence of pulmonary scarring or congestive heart disease, normal X-rays may result in poor diagnostic specificity. In some cases, X-ray CT imaging (computed tomography) and other laboratory analyses, such as sputum testing, blood cultures, and C-reactive protein (CRP) testing, may be needed to differentiate pneumonia. None of the above tests can be considered the gold standard. In hospitals in developed countries, the "reference standard" used is a clinical diagnosis supplemented by auscultation, radiology, laboratory tests, and microbiology, as needed. Often, suspicion of pneumonia is sufficient to prescribe antibiotics. Even in hospitals, it is difficult to distinguish between viral wheezing and viral pneumonia (the most common form of pneumonia).

[0006] Differential diagnosis of pneumonia and other respiratory illnesses is a complex problem worldwide. Other conditions requiring attention include bronchiolitis, asthma, viral wheezing, pneumonia, tracheobronchitis (TBM), and laryngitis—all of which occur in children. Malaria and congestive heart disease can also present with similar symptoms. Diagnosis is generally difficult even in hospital settings. Long-term surveillance in home / community settings is virtually impossible.

[0007] In resource-poor regions of the world where pneumonia is rampant, it is also difficult to find well-trained healthcare workers with specialized auscultation and clinical skills. Pneumonia management in these areas largely depends on community workers visiting remote communities.

[0008] To address these issues, the World Health Organization (WHO) has developed a simple clinical procedure for classifying pneumonia in resource-constrained areas. These classifications directly lead to interventions such as antibiotic prescriptions and hospitalization. The WHO procedure uses cough symptoms (and / or difficulty breathing) as a screening feature for pneumonia; then respiratory rate is used to determine the presence of pneumonia. If symptoms such as chest collapse are also present, the illness is further classified as severe pneumonia.

[0009] Although the WHO procedure based on respiratory rate has performed poorly, it is the primary method for determining treatment / referral decisions in remote areas where more than 90% of children have pneumonia.

[0010] In the past, some studies have explored the performance of the WHO procedure and its variants in the diagnosis of pneumonia. They reported fairly high sensitivity (69–94%), but unacceptably low specificity (67–16%). Researchers have attempted to improve the specificity of the WHO criteria using different approaches. These approaches include increasing the WHO procedure by considering fever and other symptoms of pneumonia (nasal flaring, poor sleep, chest retraction, cough lasting more than two days, etc.). These efforts have resulted in sensitivity and specificity ranging from 20–90%, but only at the cost of lower sensitivity, and vice versa. They are also affected by the fact that the higher the complexity of the measurement, the more difficult it is to train community workers to reliably implement the procedure in field visits.

[0011] The WHO procedure uses respiratory rate as a key indicator. However, obtaining respiratory rate in children is very difficult. Organizations such as the Bill & Melinda Gates Foundation, UNICEF, and the World Health Organization have invested significant resources in developing respiratory rate counters. For examples, see https: / / www.path.org / publications / files / TS_update_rr_counters.pdf. In the manual diagnosis of pneumonia, different clinical signs and measurements are dichotomized to facilitate clinician assessment. For example, a respiratory rate above the age index threshold is used to indicate the presence of pneumonia. Clinicians may also notice the presence of cough, fever, chest retractions, etc.

[0012] The problem with using only cough sounds is the number of coughs (5-10 cough events) required for reliable analysis. Infants and young children are known to be less likely to cough when needed for analysis. Furthermore, it is known that cough symptoms may disappear due to debility as the patient's condition worsens.

[0013] Therefore, it can be recognized that a key development still lacking in the global fight against pneumonia is the absence of rapid, low-cost diagnostic methods / systems [1, 8-14]. Even with state-of-the-art equipment, it is difficult to accurately and precisely diagnose each case, especially in situations with limited resources.

[0014] One object of the present invention is to provide a method and apparatus for assisting in the diagnosis of disease states such as pneumonia, which is directly usable by clinicians and solves one or more of the aforementioned problems of the prior art. Summary of the Invention

[0015] According to a first aspect of the present invention, a method is provided for automatically providing a patient's disease state diagnosis to the patient's caregiver, comprising the following steps: It provides diagnostic application software products that include multiple diagnostic models derived from studies of populations including subjects with positive and non-disease states. The operation includes the input / output interface of the electronic device's memory for storing software products, prompting caregivers to identify multiple disease diagnostic parameters and values ​​for the patient; The processor of the operating electronic device selects a diagnostic model from memory based on the identified disease diagnostic parameters; The processor is operated to apply the values ​​of disease diagnostic parameters to the selected diagnostic model; and Based on the results of applying parameters to the diagnostic model selected by the caregiver, the diagnosis is presented to the caregiver on the input / output interface for the caregiver to provide treatment to the patient.

[0016] Preferably, the diagnostic parameters include: respiratory rate, presence of fever, presence of runny nose, number of days of runny nose, number of days of cough, presence of chest retraction, temperature, BMI (body mass index), and oxygen saturation level. In one embodiment of the invention, the method includes selecting a diagnostic model by referring to one or more lookup tables that associate the diagnostic performance of the model with available diagnostic parameters.

[0017] The method may include prompting the user to select a diagnostic model optimized for “sensitivity,” “specificity,” or “accuracy,” wherein the method includes operating an electronic device to select one of the diagnostic models while taking into account the optimization options.

[0018] In one implementation, the method includes operating the device to determine whether the value of a diagnostic parameter indicates a dangerous sign in the patient.

[0019] This method may include checking whether a patient’s diagnostic values, according to World Health Guidelines, indicate that the patient presents with general signs of danger.

[0020] This method may include saving diagnostic results to a remote server, thereby saving the diagnostic results and comparing them with multiple diagnostic devices.

[0021] In some embodiments of the invention, the method includes prompting the recording of at least one cough sound from the patient.

[0022] This method may include requiring the recording of no more than two coughs from the patient.

[0023] Preferably, the method includes applying at least one cough sound from a patient to a cough feature extraction engine of a diagnostic application to generate cough features.

[0024] Patient cough characteristics can be applied to diagnostic models to aid in diagnosis.

[0025] In a preferred embodiment of the invention, the diagnostic parameters include respiratory rate, temperature, heart rate, and cough sound analysis.

[0026] According to another aspect of the present invention, a diagnostic device is provided, configured to prompt a clinician to input diagnostic information of a patient, and further configured to automatically provide a diagnosis to the clinician, the diagnostic device comprising: The electronic storage device for storing instructions includes a diagnostic application software product, which includes multiple diagnostic models derived from studies of pneumonia-positive and non-pneumonia-positive subjects. An electronic processor that communicates with electronic memory and executes instructions; A user interface that responds to the electronic processor for processor prompts and receiving diagnostic information; The processor is configured by the diagnostic application in use, and the user interface presents the patient's diagnosis by applying diagnostic information to at least one of the diagnostic models.

[0027] Preferably, the diagnostic device includes a microphone and an audio interface for connecting the microphone to an electronic processor.

[0028] Diagnostic applications may include a cough feature extraction engine, and the processor in use may apply cough sounds to the cough feature extraction engine to generate its cough features.

[0029] In one embodiment of the invention, the diagnostic device is configured by a diagnostic application in use to apply cough features to at least one of the diagnostic models.

[0030] According to another aspect of the present invention, a method for automatically diagnosing pneumonia in a patient is provided, comprising: Use an input / output interface device to obtain the values ​​of two or more diagnostic parameters of the patient from the patient's caregiver; The diagnostic parameters include: respiratory rate, presence of fever, presence of runny nose, number of days of runny nose, number of days of cough, presence of chest retraction, body temperature, BMI (body mass index), and oxygen saturation level. Using a processor device operably connected to an input / output interface, two or more diagnostic parameters are applied to an electronic memory storing multiple pre-compiled pneumonia diagnostic models, thereby identifying the optimal diagnostic model among the multiple models; Apply the values ​​of two or more diagnostic signs to the identified best diagnostic model to generate diagnostic output from the diagnostic model; and The diagnostic output is used to operate the input / output interface device to indicate to caregivers whether or not pneumonia is present in the patient, so that caregivers can provide care to the patient. The pneumonia diagnostic model was derived from studies on pneumonia-positive patients and non-pneumonia-positive subjects.

[0031] According to another aspect of the present invention, a method for operating an electronic processor-based electronic device to diagnose a patient's disease state is provided, comprising the following steps: Install diagnostic application software products on electronic devices; the diagnostic applications include a variety of diagnostic models. The prompt identifies multiple disease diagnostic parameters and values ​​for the patient; Operate electronic devices to select one of the diagnostic models based on the identified disease diagnostic parameters; Operate electronic devices to apply values ​​of disease diagnostic parameters to a selected diagnostic model; and The diagnosis is presented to clinicians based on the results of applying parameters to the selected diagnostic model. Attached Figure Description

[0032] Preferred features, embodiments, and variations of the present invention will become apparent from the following detailed description, which provides sufficient information for those skilled in the art to practice the invention. This detailed description should not be construed as limiting the scope of the foregoing summary of the invention in any way. The detailed description will be referenced to the following figures: Figure 1 This is a block diagram of a diagnostic device in use according to an embodiment of the present invention.

[0033] Figure 2 It is based on the embodiments of the present invention. Figure 1 The flowchart shows the method performed by the diagnostic equipment.

[0034] Figures 3 to 5 The image depicts a screen generated by a diagnostic device for detecting WHO general risk signs in patients.

[0035] Figure 6 The image depicts a screen generated by a diagnostic device, used by clinicians to input patient diagnostic information.

[0036] Figure 7 The image depicts a screen generated by a diagnostic device to confirm the entered diagnostic information with the clinician.

[0037] Figure 8 It is shown Figure 2 The flowchart at box 43 is the flowchart of the diagnostic model selection process for the diagnostic equipment.

[0038] Figure 9 This is a flowchart illustrating the process of selecting the best diagnostic model for the diagnostic equipment.

[0039] Figure 10 The screen depicted shows the diagnostic equipment generating information while analyzing diagnostic data.

[0040] Figure 11 It depicts a screen generated by diagnostic equipment, which presents the analysis results to clinicians.

[0041] Figure 11A This is a block diagram of a diagnostic device according to another embodiment of the present invention, the diagnostic device being configured to take into account a patient's cough sounds.

[0042] Figures 11A to 11D It is by Figure 11A The diagnostic equipment generates a screen that is used to alert the patient to their cough sounds and to analyze and present the patient's diagnostic results.

[0043] Figure 12 This is a flowchart detailing the process of analyzing cough sound data to design a diagnostic model for embodiments of the present invention.

[0044] Figure 13 This is a flowchart illustrating a method for training a diagnostic model used in embodiments of the present invention.

[0045] Figure 14 includes a graph showing the average Sn (sensitivity) and Sp (specificity) values ​​of the trained models calculated from the 2-11 month age group. Frames from top to bottom are arranged according to the number of features used to create the model: one feature acquired at a time (top), two features (middle), and three features (bottom). n is the number of possible feature combinations using one, two, or three features at a time. A significant improvement in specificity can be seen as more features are used.

[0046] Figure 15 ROC curve analysis was included for the 2-11 month age group. The solid line in each frame represents the average ROC curve formed over k iterations. The cross on each frame represents the SD of Sn and 1-Sp at the indicated point. On each frame, the performance of the WHO / IMCI procedure in resource-constrained areas is illustrated (see boxes). The center of the box represents the average performance, and the height and width of the box represent the SD.

[0047] Figure 16 ROC curve analysis was included for the 12-60 month age group. The solid line in each frame represents the average ROC curve formed over k iterations. The cross on each frame represents the SD of Sn and 1-Sp at the indicated point. On each frame, the performance of the WHO / IMCI procedure in resource-constrained areas is illustrated (see boxes). The center of the box represents the average performance, and the height and width of the box represent the SD.

[0048] Figure 17 This graph shows the training and testing performance of two optimal LRM models for the 2-11 month and 12-60 month age groups. The optimal LRM model for the 2-11 month group is the model using the runny nose / runny nose days / respiratory rate / temperature and runny nose / runny nose days / respiratory rate / heart rate characteristics. As for the 12-60 month group, the best model is based on the number of days of fever / cough / heart rate / chest retraction and the number of days of runny nose / runny nose days / respiratory rate / temperature characteristics. Detailed Implementation

[0049] First refer to Figure 1 The diagram illustrates a diagnostic device 1 according to an embodiment of the present invention. The diagnostic device 1 comprises a unique combination of hardware and software specifically designed to assist clinicians in diagnosing and thus treating disease conditions such as pneumonia. Figure 1In the illustrated embodiment of the invention, device 1 includes one or more processors in the form of microprocessors 3 and an electronic memory 5 accessible to the microprocessor 3, which stores commands executable by the microprocessor 3 to cause diagnostic device 1 to process the diagnostic signs of patient 2 and provide a diagnosis to a nurse in the form of a clinician 4. Diagnostic application 6 includes multiple diagnostic models 20 and a model lookup table 22 stored in memory 5. Microprocessor 3 uses model lookup table 22 to determine the most suitable diagnostic model based on available diagnostic parameters of patient 2. As described in Appendix B of this specification, referring to Table 9, the inventors list their quantitative studies of over 131,000 combinations of pneumonia-related features to obtain the diagnostic models used.

[0050] Electronic memory 5 includes operating system 8, for example Android Operating system or Apple iOS An operating system is provided for execution by the microprocessor 3. The electronic memory 5 also includes diagnostic application software 6 according to a preferred embodiment of the invention.

[0051] Microprocessor 3 communicates with multiple peripheral components 9 to 23 via data bus 7, such as Figure 1 As shown. The data bus consists of metal rails that transmit electrical data signals 13 between various components of device 1. Diagnostic device 1 is capable of establishing voice and data communication with voice and / or data communication network 31 via WAN / WLAN component 23 and radio frequency antenna 29.

[0052] although Figure 1 The diagnostic device 1 shown is microprocessor-based, but it will be appreciated that it can also be implemented using other types of technologies. For example, it can be implemented using a field-programmable gate array (FPGA) or discrete logic circuits.

[0053] Diagnostic device 1 does not require external physiological sensors and operates through physical contact with patient 2 or access to communication network 31. As will be explained, the method is automated and easy to use. In some embodiments, device 1 collects patient metadata, diagnostic and treatment / referral information (e.g., whether antibiotics were administered), and GPS coordinates of the diagnostic location. The device records the information or transmits it via WAN data network 31, such as the Internet, to cloud server 33 when / if network 31 is available. In one embodiment, the device enables the tracking of epidemics. It can also consider the progression of an epidemic at the time of patient diagnosis by accessing geographic information stored in cloud server 33 used for disease diagnosis in other locations using similar devices.

[0054] Now for reference Figure 2 The diagram shows a flowchart of a method according to a preferred embodiment of the present invention, wherein the diagnostic device is implemented under the control of instructions encoded into a diagnostic application 6.

[0055] exist Figure 2 At box 35, the microprocessor 3 operates the input / output interface in the form of an LCD touch screen interface 11 to display prompts to users such as clinicians to input diagnostic parameters for patient 2. Figure 3 The diagnostic device 1 displays a data input screen 57 on an LCD touchscreen interface 11 for the clinician 4 to input patient parameters.

[0056] Diagnostic procedure 6 includes instructions based on the inventor's findings, namely that at least 17 patient parameters can be used to diagnose disease states such as pneumonia, as shown in Table 1A below:

[0057] Table 1A - 17 Diagnostic Parameters for Pneumonia The inventors have discovered that the most direct patient parameters and effective ways to determine the presence of respiratory disease states such as pneumonia are the nine diagnostic parameters listed in Table 1B below.

[0058]

[0059] Table 1B - A subset of 9 key patient diagnostic signs and parameters A preferred embodiment of diagnostic device 1 is configured to operate within the framework of the World Health Organization (WHO), by providing screen 57 for clinicians 4 to select WHO screening criteria, namely “cough” and “dyspnea,” and also accessing the WHO general risk signs listed on screen 57. Figure 2 In the flowchart at box 37, if clinician 4 has already indicated on screen 57 that patient 2 is displaying WHO general risk signs, then the microprocessor control switches to box 51, and microprocessor 3 operates screen 59. Figure 4 This alerts clinicians 4 to patient 2 indicating severe pneumonia or illness according to WHO general risk signs criteria. Alternatively, if the patient's parameters at box 37 do not indicate WHO general risk signs, control redirects to box 39. At box 39, the microprocessor checks whether the patient's parameters indicate that WHO criteria for further screening have been met. If they are not met, control redirects to box 53, at which point the microprocessor 5... Figure 5 The screen 61 shown in the figure instructs the clinician 4 to display neither WHO general risk signs nor WHO screening criteria for the patient 2.

[0060] Alternatively, if at box 39, the microprocessor 5, executing further instructions of the diagnostic application 6, determines that the patient's parameter indications meet the WHO screening criteria parameters, then control turns to box 41.

[0061] At frame 41, microprocessor 5 enables screen 63 ( Figure 6The diagnostic parameters are displayed on the LCD touchscreen 11 to prompt the clinician 4 for the patient.

[0062] Screen 63 displays selection buttons 65A, ..., 65I for the nine patient parameters previously referenced in Table 1B.

[0063] In the currently described implementation, the clinician 4 needs to use selection buttons 65A, ..., 65I to select at least three diagnostic parameters from Table 1B. The control then navigates to box 43 ( Figure 2 ), where microprocessor 3 displays screen 67 on interface 11 ( Figure 7 In the preferred embodiment of the invention described herein, screen 67 includes selection buttons 69, 71, and 73 for the clinician 4 to indicate a preference for a diagnostic model, used to optimize the diagnosis's "sensitivity," "specificity," or "accuracy." These terms have the following meanings: Sensitivity (also known as true positive rate, recall rate, or detection probability in some areas) measures the proportion of positive cases that are correctly identified (e.g., the percentage of patients correctly identified as having the condition).

[0064] Specificity (also known as the true negative rate) measures the proportion of correctly identified negative cases (e.g., the percentage of healthy people correctly identified as not having the condition).

[0065] Accuracy (ACC) = (ΣTrue Positives + ΣTrue Negatives) / ΣTotal Number of People In box 43, microprocessor 3 selects the optimal diagnostic model from models 20 stored in memory 5 based on the submitted patient parameters. To select the optimal model from models 20 based on the information entered by the clinician, microprocessor 3 (as configured in diagnostic procedure 6) queries diagnostic model lookup table 22 stored in memory 5 and generated by the inventors. (It will be appreciated that in other embodiments of the invention, where a high-power processor is used, the required information can be generated on demand without resorting to a lookup table.) The contents of Diagnostic Lookup Table 22 are as follows: Table 2A - Diagnostic models for patients aged 2 to 11 months, ranked according to reduced sensitivity.

[0066] Table 2B-2 Diagnostic models for patients aged 2 to 11 months based on reduced specificity ranking Table 2C-2 Diagnostic model for patients aged 11 months to 20 years based on reduced specificity ranking.

[0067] Table 3A - Diagnostic models for patients aged 12 to 60 months, ranked according to reduced sensitivity. Table 3B - Diagnostic model for patients aged 12 to 60 months based on reduced specificity ranking.

[0068] Table 3C - Diagnostic models for patients aged 12 to 60 months, ranked according to reduced accuracy.

[0069] The patient diagnostic signs parameters that constitute each combination in the combination columns of Tables 2A to C below have previously been defined in Table 1B.

[0070]

[0071]

[0072] Table 2A (above) - Diagnostic models for patients aged 2 to 11 months, sorted by reduced sensitivity.

[0073]

[0074]

[0075] Table 2B (and above) - Patients aged 2 to 11 months - Diagnostic models based on reduced specificity ranking

[0076]

[0077] Table 2C (above) - Patients aged 2 to 11 months - Diagnostic model based on reduced specificity ranking.

[0078]

[0079]

[0080]

[0081] Table 3A (and above) - Diagnostic models for patients aged 12 to 60 months, sorted by reduced sensitivity.

[0082]

[0083]

[0084]

[0085] Table 3B (above) - Patients aged 12 to 60 months - Diagnostic model based on reduced specificity ranking.

[0086]

[0087]

[0088]

[0089] Table 3C (and above) - Diagnostic models ranked according to reduced accuracy for patients aged 12 to 60 months.

[0090] As will be noted in the table above, a maximum of four diagnostic parameters are used in combination. This is because the inventors' research surprisingly showed that using combinations of more than four diagnostic features increases computational complexity without providing any substantial improvement in diagnostic performance.

[0091] If the clinician selects "Accuracy" optimization, the optimal model will be the top row of Table A3 (if the patient is 2 to 11 months old) and the top row of Table B3 (if the patient is 12 to 60 months old). However, if the parameters entered by the clinician do not allow the use of the top row #1, the microprocessor will check if row #2, and so on, can be used. Since the listed parameters (among the nine parameters listed in Table 1B) are common measurements, it is usually not necessary to move down the rows to find the first usable model.

[0092] For example, Table 2B: 2-11 months: Parameters: #1 ACDI (Respiratory Rate, Runny Nose, #Number of Days with Runny Nose, Oxygen Saturation) >> Sens 91%, spec 70%.

[0093] #13 ABCD (Respiratory rate, fever, runny nose, #number of days with runny nose) >>sens 91%, spec 61%, etc.

[0094] Figure 8 The flowchart details how the microprocessor 3, under the control of the diagnostic application 6, performs... Figure 2 The model selection process is performed within box 43 of the flowchart. Using data from tables A1-A3 or B1-B3 (depending on age), the clinician's preference for optimizing sensitivity, specificity, or accuracy, and input patient parameters, device 1 uses... Figure 9 The process shown automatically selects a diagnostic model with a specific number.

[0095] In box 45 ( Figure 2 The patient values ​​input for each selected diagnostic parameter are applied by the microprocessor 3 to the selected optimal diagnostic model. Simultaneously, this process is displayed on the LCD touchscreen interface 11. Figure 10 The screen depicted in the image.

[0096] At box 47, the microprocessor 3, operating under the control of the diagnostic application 6, operates the LCD touchscreen interface 11 to present the diagnostic results to the caregiver, such as... Figure 11 As shown on the screen.

[0097] Return to Figure 2 In the flowchart, at box 49, the microprocessor 3 can locally save the diagnostic results to memory 5 or externally save them to cloud server 33 via data network 31 through WAN / WLAN interface and antenna 29. At box 50, if the diagnostic result indicates a disease state such as pneumonia, then the caregiver can use this information, as shown in box 52, to apply treatment to patient 2, for example, by organizing hospitalization and / or providing antibiotics or antiviral drugs. Then, at box 55, the clinician 4 can indicate the desire to diagnose another patient, in which case the process is repeated for that other patient.

[0098] As will be explained in the appendix to this specification, the inventors developed diagnostic models from quantitative test data from a subject population including patients with respiratory diseases. Various models were configured to classify patients as “affected” or “non-affected” based on diagnostic parameters input by clinicians. A preferred method for developing these models is using a logistic regression machine (LRM). Other types of classification decision engines, such as support vector machines (SVM), may also be used.

[0099] As explained in Section 1.2.2 of Appendix A of this document, the inventors have discovered that using only the cough sound characteristics of one or two coughs from a patient can provide a significant improvement in diagnostic performance. Where cough sounds from patient 2 can be recorded via microphone 25, these sounds can be processed through cough feature extraction and processed along with other diagnostic parameters in a suitable diagnostic model 20.

[0100] Now refer to Figure 11A This illustration shows another embodiment of the invention including a diagnostic device 1a, which, in terms of hardware, is identical to the previously described diagnostic device 1, but is programmed with a variant of the previously discussed diagnostic application 6a. The diagnostic application 6a includes a cough feature extraction engine 24. The cough feature extraction engine 24 includes instructions for a microprocessor 3 to analyze cough sounds received from a patient 2 via a microphone 25 and an audio interface 21, and to generate corresponding cough features. Cough feature extraction is known in the prior art and is described in International Patent Publication No. WO2013 / 142908 by the inventors Abeyratne et al., the disclosure of which is incorporated herein by reference in its entirety.

[0101] Microprocessor 3 applies cough features to diagnostic model 20, which forms part of diagnostic application 6a and is configured to consider other diagnostic parameter values ​​input by clinician 4. This is accomplished by adding cough features as grouping features to an existing lookup table, as shown in Tables 2A-3C. As previously described, the modified table, including cough sounds, is arranged along with the model in a reduced performance order. Users, such as clinician 4, can select cough sounds as a measurement if they are able to collect sounds from patient 2. Otherwise, they will not select that option, as is the case with other clinical signs. Then, following the same logic previously discussed with respect to Tables 2A-3C, diagnostic device 1 is programmed via diagnostic application 6 to select the best model corresponding to the available measurements.

[0102] Figure 11B and 11C Screens 54 and 56, generated by the microprocessor 3 on the LCD touchscreen interface 11 under the control of the diagnostic application 6, are depicted to prompt the clinician 4 to obtain at least two cough sound samples from the patient 2. Figure 11D A screen 57 generated by the microprocessor 3 on interface 11 is depicted for providing analysis results to the clinician 4. In other embodiments, as discussed in Appendix A, it has been found that a single cough sound can help improve the performance of machine diagnostics.

[0103] Appendix A Diagnosis of pneumonia in children using clinical and cough characteristics 1.1 Materials and Methods 1.1.1 Research Group A total of 222 children were recruited during data collection: 93 females and 129 males, with a median age of 9 months and an interquartile range (IQR) of 4.25–20 months. Due to the lack of one or more required parameters, we excluded 123 patients for further consideration, leaving 99 children with complete data. As shown in Table 4.1, the distribution of these 99 children is closely representative of the initially recruited 222 children. There were 52 children aged 2–11 months and 47 children aged 12–60 months. Of the 99 children, 67 had positive pneumonia, while the remaining 32 had asthma, bronchitis, bronchiolitis, heart disease, malnutrition, wheezing, etc. The non-pneumonia patients served as the control group for this study.

[0104] Table 4.1 Representative children between the initial recruitment and the actual number in the study

[0105] 1.1.2 Data Analysis Figure 12Flowchart 1200, presented in the document, details the process used to analyze our data. For the analysis, the dataset was divided into two age groups: 2–11 months and 12–60 months. This followed the WHO / IMCI procedure, which applied different respiratory rate thresholds to each age group. Clinical and cough characteristics for each group were processed by inputting a feature matrix.

[0106] The LRM classifier is used to train a model for each feature combination. This technique creates a non-linear model based on feature parameters and applies different weights to each parameter. The LRM training process adjusts the weights so that each model is suitable for distinguishing between pneumonia patients and non-pneumonia patients. We analyzed feature combinations of one to six features in detail. The process was divided into several stages: clinical features only, clinical features plus cough features from one cough, and clinical features plus cough features from two coughs. Before analysis, cough features from multiple coughs were averaged into a separate subset.

[0107] Leave-one-out (LOO) cross-validation is used to validate the results. In each iteration, one patient is designated as the LOO patient, and the remaining patients become part of the training set. An LRM model is created based on the training set and a cutoff threshold, with careful selection of S... n ≥90% and S p The higher the value, the better. Then, the model parameters are fixed and used to evaluate the training set and LOO individuals. The trained model is applied to the training set to provide training performance, enabling the classification of LOO patients using the trained model. In the next iteration, another person is assigned as LOO, while the others become part of the training set. At the end of the iteration process, each patient is assigned as a LOO individual once, and the program performance is calculated.

[0108] 1.2 Results In this section, we will present our analysis results in three phases. First, we compare the WHO / IMCI program performance for each age group with clinical features only. Then, we compare it with the clinical features plus one cough feature. Finally, we compare the WHO / IMCI program results with the program using the clinical features plus two cough features.

[0109] 1.2.1 Clinical features only The results in Table 4.2 show the WHO / IMCI classification and our procedure performance using only clinical characteristics for each age group. For the 2–11 month age group, Sn and Sp for the WHO / IMCI procedure were 91.2% and 26.7%, respectively. In the older age group, the corresponding Sn and Sp were 83.3% and 11.8%, respectively. The best model in the 2–11 month age group using only clinical characteristics showed that, using fever, number of days of cough, temperature, and heart rate, S n and S pThe rates were 86.5% and 60.0%, respectively. In the 12-60 month age group, the optimal model was determined using fever, body temperature, runny nose, and heart rate, S... n and S p The figures were 90.0% and 76.6%, respectively.

[0110] 1.2.2 Clinically, add one cough. When the characteristic of only one cough was included in the mixture, there was a significant improvement in performance. For the 2-11 month age group, the best model showed S n and S p They were 89.2% and 73.3% respectively. The S p The value is equivalent to 275% of the WHO / IMCI procedure, while S n Only a small decrease. Values ​​are shown in Table 4.3. For the 12–60 month age group, the optimal model using clinical characteristics and a single cough yielded 90% and 82.4% S, respectively. n and S p Conducted. Compared with WHO / IMCI performance, S p This is equivalent to 82.4 / 11.8 = 698%, while S n Slightly higher than WHO / IMCI.

[0111]

[0112]

[0113]

[0114] 1.2.3 Clinically, two additional coughs were observed. Table 4.4 shows the LRM performance when clinical features are combined with the characteristics of the two coughs. The optimal model for the 2–11 month age group utilizes the number of cough days, respiratory rate, and temperature, as well as cough characteristics. This leads to S n and S p They were 89.2% and 86.7% respectively. p This is equivalent to 325% of the WHO / IMCI procedure. In the 12-60 month age group, the best LRM model used respiratory rate, temperature, heart rate, and cough characteristics. The model showed 90.0% Sn and 88.2% S. p In this case, S p This is equivalent to 747% of the corresponding portion of the WHO / IMCI.

[0115] 1.3 Discussion The aim of this study was to investigate the use of combining commonly used clinical parameters with cough sound analysis in the diagnosis of pneumonia in children. A key requirement for this goal was to outperform the WHO / IMCI procedure. Our results indicate that incorporating cough sound analysis does indeed help in identifying pneumonia in children compared to relying solely on clinical parameters. In this study, the WHO / IMCI procedure's S... n and S p The ranges were 83-91% and 12-27%, respectively. Our best-performing model, which had two cough sounds, achieved S0.05 at 89-90% and 87-88%, respectively. n and S p The range is used to classify pneumonia. These are the results of LOO validation.

[0116] Coughing is one of the most common symptoms in children with respiratory illnesses. The WHO / IMCI program takes into account the presence of coughing but does not fully utilize it. Our method extracts key features from recordings of cough sounds and uses them to enhance our clinically-based diagnostic features for pneumonia. We have demonstrated that this provides optimal results compared to using only cough sounds or clinical features.

[0117] Despite recent technological advancements, the field of cough sound analysis remains relatively unchanged. We believe that cough contains more information about lung physiology than currently known. It is hypothesized that diseased lungs may alter lung physical characteristics and the characteristic sound of cough in disease-specific ways. In the case of childhood pneumonia, our previous research has shown that cough is a feasible parameter for diagnosis. [7,62] The number of coughs used there was 15 per person, far higher than the number used in this paper. We have now demonstrated that, combined with clinical parameters, fewer coughs may be needed to produce an accurate diagnosis.

[0118] 1.4 Conclusion We have demonstrated the potential benefits of combining clinical and cough characteristics for the diagnosis of pneumonia in children in resource-poor areas. Our developed model showed approximately 90% sensitivity and specificity, ranging from 325 to 747% higher than the WHO / IMCI procedure. The most effective key parameters were a combination of respiratory rate, temperature, heart rate, and cough sound analysis. Clinical characteristics are readily measurable, and in the absence of key parameters, we can switch to other models that still perform well.

[0119] It should be noted that this study is currently based on a relatively small number of participants (n = 99) and no gold standard for the diagnosis of pneumonia has been established. The reference standards we used are derived from clinical diagnosis by pediatricians using auscultation, laboratory analysis, chest X-ray (if applicable), and the participants' treatment response to the clinical course of the disease.

[0120] -------------End of Appendix A------------ Appendix B. Detailed mathematical analysis of simple clinical measurements In 1990, the World Health Organization (WHO) and UNICEF proposed the WHO classification of pneumonia in children in resource-poor areas. This is the diagnostic method currently used by community health workers in resource-limited environments as a rapid, low-cost alternative in frontline health facilities. Table 5 shows the WHO / IMCI classification guidelines for pneumonia in resource-poor areas.

[0121] Table 5 - WHO / IMCI guidelines for pneumonia classification in resource-poor regions [35,36]

[0122] The WHO / IMCI guidelines state that if a patient develops symptoms of cough / dyspnea, further screening should be conducted.

[0123] Consider the respiratory rate, and if it exceeds the limits (50 breaths per minute (bpm) for children aged 2 to 11 months, and 40 bpm for children aged 12 to 60 months), then non-severe pneumonia is declared. Warning signs such as low chest retraction and inability to eat or drink will place the patient in the severe pneumonia category and require immediate attention.

[0124] Researchers generally recognize the limitations of WHO standards, which are sensitive but not highly specific. [9,15,16] Over the years, others have suggested increasing fever.

[17] snoring and snot [8] Temperature and oxygen saturation

[18] Rambaud-Althaus et al. proposed a decision tree-based approach to combining signs to improve the accuracy of clinical diagnosis. [8] Researchers at the Childhood Health Pneumonia Etiology Study (PERCH) developed their own criteria for interpreting symptoms and signs based on the WHO's clinical case definition criteria for pneumonia.

[16] .

[0125] All of these approaches have made significant contributions to addressing the global burden of pneumonia, but are subject to similar limitations to World Health Organization standards, primarily affecting resource-poor regions. These approaches also rely on health workers using basic binary decision-making around fixed thresholds to perform measurement and interpret data.

[0126] method Research Organization The clinical data used in this study were collected in collaboration between Gazamada University-Sardjito Hospital in Yogyakarta, Indonesia, and the University of Queensland in Brisbane, Australia. Data collection began in December 2010 and continued until March 2014. The study protocol was approved by the ethics committees of Sardjito Hospital and the University of Queensland. Inclusion / exclusion criteria are shown in Table 6 (Supplementary) below: Table 6 - Patient Recruitment Criteria

[0127] Patients are included if they present with any two of the following symptoms: cough, sputum, increased wheezing, and temperature >37.5°C. Parental consent is sought prior to inclusion if the patient meets the criteria; otherwise, the patient is excluded. Exclusion criteria also apply to patients presenting with symptoms of advanced disease, advanced lung cancer, and / or requiring intranasal IV infusion, as these may distort the results. As a precaution, patients presenting with infusion spread disease are also excluded.

[0128] Diagnostic definition The reference diagnosis used in this study was an overall diagnosis provided by a pediatrician based on clinical presentation, laboratory tests, chest X-ray, and the clinical course of the disease. X-rays were performed only on subjects clinically suspected of having pneumonia, and only in other cases where definitive diagnosis was required.

[0129] Therefore, not all participants underwent a chest X-ray.

[0130] Research Plan All children meeting the inclusion criteria were invited to participate in the study. Each child's medical history and clinical measurements were recorded as part of the hospitalization process. Diagnoses were obtained during the study, along with all test results collected from participants during the normal diagnosis / management of the disease. Table 7 (Supplementary) lists some of the recorded information used in the analysis in this paper.

[0131] Table 7 - List of clinical parameters recorded upon admission

[0132] Test parameters include the presence of fever, cough, dyspnea, runny nose, and chest retractions, as a binary yes / no observation. It also includes the following data as numbers: age, weight, height, respiratory rate, body temperature, BMI, oxygen saturation, and the number of days with fever, cough, dyspnea, and runny nose. Other diagnostic measures, such as blood / sputum analysis and chest X-ray, are performed only if deemed necessary by the attending physician.

[0133] research group We recruited 222 children: 93 females and 129 males, with a median age of 9 months and an interquartile range (IQR) of 4.25–20 months. Our cohort consisted of children from inpatient wards. Our aim was to identify clinically useful parameters for the diagnosis of pneumonia in resource-poor areas. The dataset contained a complete list of 134 parameters with the previously specified parameters. We excluded 88 patients from further consideration due to the absence of one or more required parameters. As shown in Table 8, the distribution of the selected 134 children closely represents the initially recruited 222 children.

[0134] Table 8 - Representative Children Between Initial Recruitment and Actual Number of Students in the Study

[0135] There were 71 children in the 2-11 month age group and 63 children in the 12-60 month age group. Of the 134 children, 96 were diagnosed with pneumonia, while the remaining 38 children had a combination of asthma, bronchitis, bronchiolitis, heart disease, malnutrition, wheezing, etc.

[0136] The non-pneumonia group served as the control group in this study.

[0137] Analyze data Figure 13 The flowchart 1300 presented in the document details the process used to analyze our data. The dataset is divided into two age groups: 2-11 months and 12-60 months.

[0138] The clinical characteristics of each group are input into the feature matrix for processing.

[0139] Using k-fold cross-validation, each age group is randomly divided into k replicates. An iterative process is then employed, where one set of data is reserved as the test set, and the remaining data is used to train a logistic regression model (LRM). A general explanation of the logistic regression method used in medical applications can be found in KL Sainani.

[26] The paper.

[0140] LRM outputs the probability of pneumonia presence based on specified predictors, and applies a cutoff threshold to this probability to formulate an output binary decision. This threshold is carefully selected after receiver operating characteristic (ROC) analysis to separate positive and negative pneumonia cases as cleanly as possible.

[0141] In the LRM design, we begin by using one feature at a time and calculating the performance of the resulting model. Then, we exhaustively search for all combinations considering two features at a time. This process continues until we consider all 17 features at once. In each iteration, the sensitivity (S) of the trained model is adjusted. n ), specificity (S p ), accuracy (Acc The training model is evaluated using the area under the curve (AUC). AUC can only be used on training data to set diagnostic thresholds.

[0142] Please note that in each iteration of k-fold cross-validation, the dataset is split into non-overlapping training and test sets, and performance is evaluated separately for each set. Each iteration produces k ROC curves and a set of k training and test performance measures. Each iteration also generates k trained LRM models. The trained LRM models are used to compute the performance on the training set. The LRM models are then fixed and used on the test dataset to compute test performance and validate the trained models. Each set of LRM models is considered final in its respective fold. Therefore, no AUC data is available for testing performance.

[0143] The best-performing model is selected based on the mean and standard deviation (SD) of its training and testing performance. These values ​​are calculated and reported in the results section. Similarly, WHO standard performance numbers for the test set are calculated and expressed using their mean and SD. We then use these performance values ​​to determine which parameters and combinations provide the best diagnostic results for the test set.

[0144] The process was iterated for each feature combination used. First, we analyzed the LRM performance using one, two, and three features at a time. Next, we thoroughly analyzed all possible feature combinations, using up to 17 features at a time. Table 9 shows the possible combinations of the number of features used when creating the LRM model.

[0145] Table 9-17 shows the total number of possible feature combinations among the available features.

[0146]

[0147] Using one feature at a time provides 17 possible combinations, while using all 17 features at once yields only one possible combination. The number of possible combinations increases significantly with the latter approach. For example, using eight features at once produces 24,310 combinations. In total, the number of models tested in this study included 131,071 combinations. Given the large number of models tested, ROC curve analysis to find the optimal cutoff threshold for each model becomes crucial. We selected ROC curve analysis targeting S... n ≥90% and S p A threshold as high as possible. This also has the benefit of reducing the false detection rate (FDR). As we mentioned earlier, our goal is to improve the S of WHO procedures. p At the same time, maintain high S n .

[0148] result In this section, we will present our analysis results, starting with the cross-validation process and the performance of the WHO / IMCI procedure in our patient groups. We then describe the model's performance and compare it with WHO results for the 2–11 month age group, followed by the 12–60 month age group.

[0149] Cross-validation technology As described in the Methods section, we use k-fold cross-validation to train and evaluate our classifier model. In this study, we set k = 8, resulting in 8-9 children per fold. Higher k values, such as the more commonly used k = 10, would result in 6-7 children per fold. We considered this insufficient and decided that k = 8 better balanced the test data. Note that in each fold of cross-validation, the training and test sets are mutually exclusive; that is, the training and test sets do not overlap.

[0150] WHO / IMCI performance We applied the WHO criteria (see Table I) to the data in each fold of the k-fold cross-validation dataset and calculated the mean and standard deviation (SD) for all folds. The results are shown in Table 10.

[0151] Table 10 - WHO standards applied to k-weighted test sets (expressed as mean + / - standard deviation)

[0152] As expected, WHO criteria yielded higher S values ​​in both age groups. n And relatively small SD, but specific S p Difference.

[0153] Our goal is to maintain the high sensitivity of the WHO procedure while improving its specificity. We will now describe the performance of the proposed method. Since the analysis was performed separately for each age group, we will first present the results for the 2–11 month age group.

[0154] Performance of the 2-11 month age group Figures 14A to 14C The figure shows the average S for models using one, two, and three features at a time for the 2-11 month age group. n and S p Values. Our feature set consists of 17 observations / measurements listed in Table 11. The total number of combinations considering one feature at a time is 17, resulting in 17 LRM models, one of which is used as input ( Figure 14A Similarly, two features at a time and three features at a time yield 136 and 680 LRM models respectively. Figure 14B and Figure 14C(Bottom frame). Overall, significant benefits were found when combining up to four features at a time in an LRM.

[0155] Figure 15 The ROC curve analysis for the 12 trained LRM models we selected for further consideration is shown. Two models were derived from single and two feature combinations, five models from three feature combinations, and three models from four feature combinations. The solid line in each frame represents the average ROC curve formed at k folds. The cross on each frame represents the S at the indicated point. n and 1-S p SD. In Figure 15 In each frame, we also graphically illustrate (see boxes) the potential performance of the WHO / IMCI program in resource-limited areas. The center of the box represents the average performance, and the height and width of the box represent the SD (Significant Degree). Table 11 shows the training and testing performance figures for 12 models.

[0156] Table 11 - Performance comparison among various models used to diagnose pneumonia in children aged 2-11 months (AUC = area under the curve, CO = cutoff threshold).

[0157]

[0158] When using a single characteristic to create an LRM for the 2–11 month age group, the best characteristic in terms of performance is respiratory rate (S / L). n For 91% and S p (35%) and chest retraction (S) n For 98% and S p (33%). These figures are closely related to the performance of the WHO criteria, as it also relies on the same characteristics of the classification of childhood pneumonia.

[0159] Individually, respiratory rate and temperature showed the highest AUC in training performance. However, compared to the respiratory rate model with comparable values, the temperature model showed a higher SUC compared to WHO standards. n and lower S p On the test dataset, both models exhibited high S-values ​​with very small SD. n and, but S p The SD variation is so large that neither model is usable on its own. This indicates that the WHO standard remains more reliable compared to single-feature LRM models.

[0160] Using two features at once can combine certain features into S p Increase to 50%, and at the same time increase S n It remains around 90%. This is the best S-value for a single feature model. pSignificant improvements were made. The best performing models were those using respiratory rate and oxygen saturation, and those using respiratory rate and fever. The combination of the two features showed a high AUC (75-82%) during training.

[0161] We continued adding features until we found the optimal LRM feature combination. On the three-feature model, the overall test performance was higher than the two-feature model. Average S n The level remains constant at around 90%, with an average S p The average level is 30% higher than the dual-feature model. n The SD level remained unchanged, but S p It decreased by 44%. The best-performing model in this category included fever, oxygen saturation, and chest retraction as parameters, S n The value was 88.1 ± 13.6%, S p The value was 61.9 ± 8.7%. This is consistent with the WHO / IMCI procedure (S...). n and S p Compared to (92.0±11.6% and 38.1±18.5% respectively), the average S n 4% lower than S p 62% higher. Compared to the WHO, S n SD is 17% higher, S p The SD is 53% smaller. Therefore, regarding S p In this regard, the best triple feature model performs far better than the WHO standard, and S n The loss was minimal.

[0162] By using four features at once, classification performance can be further improved. The best-performing model used the presence of runny nose, number of days with runny nose, respiratory rate, and temperature (91.3 ± 13.0% S). n and 70.2 ± 22.80% S p Average S n Comparable to WHO results, average S p 84% higher. However, compared to the WHO / IMCI procedure, S n and S p The SD was slightly higher. The second-best performing model used runny nose, with runny nose days, respiratory rate, and heart rate at 91.5 ± 9.2%, and S... p It is 66.0 ± 26.3%. Average S n This is also consistent with the WHO results, while S p That's 73% higher. For SD, compared to WHO, their S... n Small 20%, S p 42% higher.

[0163] Performance in the 12-60 month age group For the 12-60 month age group, the same procedure was repeated, first observing the ROC curves of the 12 trained LRM models selected for comparison, such as... Figure 16 As shown, this is the ROC curve analysis for the 12-60 month age group. The solid line in each frame represents the average ROC curve formed over k iterations. The cross on each frame represents the S value at the indicated point. n and 1-S p The performance of the WHO / IMCI procedure in resource-constrained areas is illustrated in cross-hatching on each frame. The center of the box represents the average performance, and the height and width of the box represent the SD.

[0164] Table 12 - Performance comparison between the training model and WHO standards for children aged 12-60 months (AUC = area under the curve, CO = cutoff threshold).

[0165]

[0166] The two best-performing models in the two age groups Figure 17 A comparison was made.

[0167] For individual trait categories, respiratory rate and chest retraction (individually) still showed generally best performance. The WHO / IMCI program implementation in this age group demonstrated S n It was 95.7 ± 7.6%, S p The value was 9.8 ± 13.1%. The best dual-feature LRM model in this age group also included respiratory rate as a parameter. The two best models were respiratory rate and fever versus respiratory rate and oxygen saturation.

[0168] For the triple feature, optimal testing performance was observed when using fever, temperature, and chest retraction. Regarding testing performance, the combination of S... n Reaching 92.1±15.8%, S p It reached 51.3 ± 39.0%. Average S n Still comparable to WHO standards, with a difference of 3%, but the average S of the LRM model... p 423% higher.

[0169] Compared with the WHO, S n The SD value increased by 107%, while for S p It increased by 197%. The best S in this category p It is a combination of fever, respiratory rate, and temperature, S p It was 74.8% ± 30.2%. This is higher than the WHO average S. p The increase was 663%, and the SD (symptomatic relief) increased by 130%. Compared with WHO results, the average S... nThe value also decreased by 8%, while SD was still 100% higher.

[0170] Among the models with four features, the best-performing model used the presence of fever, number of days of cough, heart rate, and chest retraction (S). n It was 94.0 ± 12.1%, S p (74.0 ± 23.3%). Average S n 2% lower than WHO's performance, average S p 655% higher. The second-best performing model used runny nose, number of days with runny nose, respiratory rate, and temperature, S n and S p The percentages were 91.4 ± 12.1% and 71.9 ± 36.4%, respectively.

[0171] Repeatability features in the best performance model As certain features recurred in the best-performing LRM across all feature combinations, we decided to systematically explore these features in order to rank the most important features among the 17 features considered.

[0172] We set a 90% threshold S for all possible combinations. n and 70% of S p From using one feature at a time to 17 features, 20 models that met the criteria were found (8 models in the 2-11 month age group and 8 models in the 12-60 year age group). Table 13 shows the number of repetitions of each feature in the first 20 feature combinations.

[0173] Table 13 shows S n >90% and S p >70% of the features appear in the model.

[0174]

[0175] Several measurements, such as respiratory rate and the observation of runny nose, are primarily present as repeatable features in a good model.

[0176] discuss A specific objective of this study was to explore whether routine clinical observation and measurement could be used to diagnose pneumonia with higher specificity than the WHO procedure while maintaining at least 90% sensitivity. Our results indicate that this is indeed possible. Our best-performing model demonstrated 91% sensitivity while achieving 70-72% accuracy based on the age of the subjects. p Compared to the WHO / IMCI procedures, these values ​​represent S p Increased by 84-655%, of which S pBetween 10% and 38%. Our results are based on k-fold cross-validation, therefore the reported results are not from the same data used to train a specific model.

[0177] The number of clinical observations and measurements required to achieve the desired performance provides useful insights for designing clinical protocols for resource-poor regions. Our results indicate that our single-feature model performs similarly to the WHO / IMCI procedure. Adding the second, third, and fourth features significantly improved performance. p And S n It remains above 90%. With more than four features, computational complexity increases without any performance improvement.

[0178] A significant contribution of this paper is the identification of the most important clinical features and measures that can significantly improve the accuracy of pneumonia diagnosis in resource-poor areas. We investigated our extensive model database to determine the clinical features and measures that meet the criteria of S n >90% and S p Features that appear repeatedly in more than 70% of the models.

[0179] Respiratory rate was a characteristic parameter in 16 models across both age groups. Oxygen saturation and chest retraction were also important parameters, appearing in 8 and 12 models out of a total of 20 models, respectively. The importance of these measurements is well-known in the medical and research communities. Our work revealed two potentially significant parameters; “presence of runny nose” and “number of days with runny nose” both appeared in 16 out of 20 models, just like respiratory rate. For the 12–60 month age group, “presence of fever” also appeared as a frequent parameter (4 out of 20 models).

[0180] Respiratory rate is the primary measurement used in the WHO / IMCI program for diagnosing pneumonia. While it appears to be an easily measurable parameter, it is difficult to achieve in resource-poor regions. Therefore, a significant portion of global pneumonia diagnostic resources has been allocated to improving techniques and protocols for measuring respiratory rate. [nn] Without reliable respiratory rate measurements, the WHO / IMCI method cannot be used in the field.

[0181] Our findings indicate that while respiratory rate is an important parameter, it is not a necessary condition for diagnosing pneumonia. For example, for the 2-11 month age group, a model using four features—age, presence of fever, presence of cough, and number of days of cough—can yield S... n = 83.5 ± 14.5% and S p = 67.3±23.9%. In another age group, the model's S n and S pThe percentages were 91.7 ± 17.8% and 51.0 ± 34.6%, respectively. In the two-characteristic model, using the combination of fever and cough days resulted in a lower S-value for the 2–11 month age group. n and S p The percentages were 90.2 ± 14.0% and 44.3 ± 25.9%, respectively. For the older age group, the model used S... n = 85.5 ± 15.2% and S p = 41.9 ± 47.7%. These results confirm our previous observation that respiratory rate may not add additional value when mathematical features derived from cough are available for diagnosing pneumonia. 31 .

[0182] Recently, interest has rekindled in using pulse oximeters to reduce childhood pneumonia mortality in resource-poor areas. [27-29] Hypoxemia is a diagnostic indicator of severe pneumonia, and rapid access to oxygen therapy, if possible, can improve prognosis. In our detailed model building process, we found that 8 out of 20 optimal models included pulse oximetry as a feature. Pulse oximetry can be a very useful feature. However, our results indicate that simpler combinations of features can be used as alternatives when a pulse oximeter is not available on-site. The features of the example are the presence and duration of nasal discharge.

[0183] The WHO / IMCI criteria for resource-poor regions were designed to be highly sensitive in detecting pneumonia (94% for patients <24 months and 62% for patients ≥24 months).

[17] This method also produces a large number of false positive results, reducing its specificity (16-20%).

[17] In our two previous studies on children, we observed a sensitivity of 83% and a specificity of 47% for the WHO / IMCI (n=91). [25,31] When used by physicians in conjunction with clinical and radiological analyses, the WHO / IMCI criteria functioned well, demonstrating a sensitivity of 77-81% and a specificity of 77-80%.

[32] These values ​​are comparable to those we obtained in this paper, and our method does not use laboratory or radiological measurements.

[0184] The low specificity of WHO criteria can lead to increased community-acquired antibiotic resistance and render antibiotics ineffective. It also wastes rare drug stocks and delays early treatment opportunities for diseases with overlapping symptoms, such as malaria. [10,33] In resource-constrained settings with only WHO / IMCI criteria, up to 30% of cases present with symptoms compatible with both malaria and pneumonia, requiring dual treatment.

[34] One of these treatment methods may be redundant. Even in the absence of key parameters such as respiratory rate, the method proposed in this paper may help address these issues by producing more accurate results.

[0185] The approach we adopted in this paper is unique. We systematically exhausted all possible feature combinations across 17 feature sets. We constructed and tested a total of 131,071 models, each using a different feature combination. In the literature, there are instances where the WHO / IMCI procedure adds one or two other carefully selected clinical features (e.g., fever, blood oxygenation) for human interpretation. For example, Cardoso et al. in their 2010 study…

[17] The original WHO / IMCI procedure was incorporated into the experimental data, showing an increase in specificity to 44% (<24 months age group) and 50% (24–60 months age group). However, the sensitivity decreased to below the level of WHO / IMCI. Specifically, in the 24–60 months age group, neither the original WHO / IMCI nor the modified method achieved a sensitivity above 62%. Our proposed method achieves a sensitivity above 90% while maintaining specificity within the 70–72% range. Without requiring human interpretation of characteristics, our method provides a decision-making apparatus.

[0186] In an independent development, Naydenova et al. published results on a method that combines several features using machine learning approaches.

[30] They reported that oxygen saturation, temperature, respiratory rate, and heart rate yielded the best performance in their model (sensitivity 96.6%, specificity 96.4%). In our work, the same combination of features resulted in much worse performance (sensitivity 88.8%, specificity 40% in the 2–11 month age group; sensitivity 82.7% and specificity 35.4% in the 12–60 month age group).

[0187] Our method and Naydenova

[30] A key difference is that they used healthy individuals as control subjects, while we used children with respiratory symptoms who met the inclusion criteria as our control subjects. Our control subjects were children who had visited the hospital seeking treatment for an illness with symptoms similar to pneumonia, but were medically diagnosed with a different disease. The research questions we explored were similar to those in Naydenova.

[30] The questions addressed in these studies are entirely different, therefore the results are not comparable. Separating healthy children from those with pneumonia is a simpler problem than identifying pneumonia subjects from a group of children with a range of respiratory illnesses.

[0188] in conclusion We have developed a method for diagnosing pneumonia using logistic regression modeling, based on a variety of clinical characteristics typically recorded by patients. Our developed LRM model retains the high sensitivity of the WHO / IMCI procedure while improving the mean specificity by 84% in the 2–11 month age group and by 655% in the 12–60 month age group.

[0189] This study is currently limited by the number of participants involved (n = 134) and the diagnostic method for pneumonia. The reference criteria used in this study were a holistic clinical diagnosis aided by auscultation, laboratory analysis, and radiography (when deemed clinically necessary by the attending physician) and the clinical course of the participants' response to treatment. Due to the need to limit radiation exposure to children, X-ray imaging was not performed on all participants in the study.

[0190] References The published content of the following documents is incorporated into this paper in its entirety through citation.

[0191] [1] Fullman N, Lim S, Dieleman J, Greenslade L, Graves C, Huynh C, etal. Pushing the Pace: Progress and Challenges in Fighting ChildhoodPneumonia. (IHME) IfHMaE, editor. Seattle, WA: IHME; 2014. [2] Liu L, Oza S, Hogan D, Perin J, Rudan I, Lawn JE, et al. Global, regional, and national causes of child mortality in 2000–13, with projectionsto inform post-2015 priorities: an updated systematic analysis. The Lancet.2015;385(9966):430-40. [3] UN. We Can End Poverty: Millenium Development Goals and Beyond2105. 2015; Available from: http: / / www.un.org / millenniumgoals / childhealth.shtml . [4] Bryce J, Black R, Victora C. Millennium Development Goals 4 and5: progress and challenges. BMC Medicine. 2013;11(1):225. [5] UN. The Millenium Development Goals Report 2014. New York: UN;2014. [6] Walker CL, Rudan I, Liu L, Nair H, Theodoratou E, Bhutta ZA, etal. Global burden of childhood pneumonia and diarrhoea. Lancet. 2013;381(9875):1405-16. [7] Madhi SA, Wals PD, Grijalva CG, Grimwood K, Grossman R, IshiwadaN, et al. The Burden of Childhood Pneumonia in the Developed World: A Reviewof the Literature. The Pediatric Infectious Disease Journal. 2013;32(3):e119-e27. [8] Rambaud-Althaus C, Althaus F, Genton B, D'Acremont V. Clinicalfeatures for diagnosis of pneumonia in children younger than 5 years: asystematic review and meta-analysis. The Lancet Infectious Diseases. 2015;15(4):439-50. [9] Qazi S, Were W. Improving diagnosis of childhood pneumonia. TheLancet Infectious Diseases. 2015;15(4):372-3.

[10] WHO. Antimicrobial resistance: global report on surveillance2014. Geneva: World Health Organization; 2014.

[11] Lynch T, Bialy L, Kellner JD, Osmond MH, Klassen TP, Durec T, etal. A Systematic Review on the Diagnosis of Pediatric Bacterial Pneumonia:When Gold Is Bronze. PloS one. 2010;5(8):e11989.

[12] Chang AB, Ooi MH, Perera D, Grimwood K. Improving the Diagnosis,Management, and Outcomes of Children with Pneumonia: Where are the Gaps?Frontiers in Pediatrics. 2013;1:29.

[13] Esposito S, Principi N. Unsolved problems in the approach topediatric community-acquired pneumonia. Curr Opin Infect Dis. 2012;25(3):286-91.

[14] Rudan I, El Arifeen S, Bhutta ZA, Black RE, Brooks A, Chan KY,et al. Setting Research Priorities to Reduce Global Mortality from ChildhoodPneumonia by 2015. PLoS Med. 2011;8(9):e1001099.

[15] WHO. Consultative meeting to review evidence and researchpriorities in the management of ARI. Geneva: World Health Organization; 2004.

[16] Scott JAG, Wonodi C, Moïsi JC, Deloria-Knoll M, DeLuca AN,Karron RA, et al. The Definition of Pneumonia, the Assessment of Severity,and Clinical Standardization in the Pneumonia Etiology Research for ChildHealth Study. Clinical Infectious Diseases: An Official Publication of theInfectious Diseases Society of America. 2012;54(Suppl 2):S109-S16.

[17] Cardoso M-RA, Nascimento-Carvalho CM, Ferrero F, Alves FtM,Cousens SN. Adding fever to WHO criteria for diagnosing pneumonia enhancesthe ability to identify pneumonia cases among wheezing children. Arch DisChild. 2011;96(1):58-61.

[18] Wingerter SL, Bachur RG, Monuteaux MC, Neuman MI. Application ofthe World Health Organization Criteria to Predict Radiographic Pneumonia in aUS-based Pediatric Emergency Department. The Pediatric Infectious DiseaseJournal. 2012;31(6):561-4.

[19] Dreiseitl S, Ohno-Machado L. Logistic regression and artificialneural network classification models: a methodology review. Journal ofBiomedical Informatics. 2002;35(5–6):352-9.

[20] Stoltzfus JC. Logistic Regression: A Brief Primer. AcademicEmergency Medicine. 2011;18(10):1099-104.

[21] Shorr AF, Zilberberg MD, Reichley R, Kan J, Hoban A, Hoffman J,et al. Readmission Following Hospitalization for Pneumonia: The Impact ofPneumonia Type and Its Implication for Hospitals. Clinical InfectiousDiseases. 2013.

[22] Aliberti S, Di Pasquale M, Zanaboni AM, Cosentini R, BrambillaAM, Seghezzi S, et al. Stratifying Risk Factors for Multidrug-ResistantPathogens in Hospitalized Patients Coming From the Community With Pneumonia.Clinical Infectious Diseases. 2012;54(4):470-8.

[23] Teka Z, Taye A, Gizaw Z. Analysis of risk factors for mortalityof in-hospital pneumonia patients inBushulo Major Health Center, Hawassa,Southern Ethiopia. Science. 2014;2(5):373-7.

[24] Swarnkar V, Abeyratne U, Chang A, Amrulloh Y, Setyati A, TriasihR. Automatic Identification of Wet and Dry Cough in Pediatric Patients withRespiratory Diseases. Ann Biomed Eng. 2013;41(5):1016-28.

[25] Kosasih K, Abeyratne UR, Swarnkar V, Triasih R. WaveletAugmented Cough Analysis for Rapid Childhood Pneumonia Diagnosis. BiomedicalEngineering, IEEE Transactions on. 2015;62(4):1185-94.

[26] Sainani KL. Logistic Regression. PM&R. 2014;6(12):1157-62.

[27] Floyd J, Wu L, Hay Burgess D, Izadnegahdar R, Mukanga D, GhaniAC. Evaluating the impact of pulse oximetry on childhood pneumonia mortalityin resource-poor settings. Nature. 2015;528(7580):S53-S9.

[28] Ginsburg AS, Delarosa J, Brunette W, Levari S, Sundt M, LarsonC, et al. mPneumonia: Development of an Innovative mHealth Application forDiagnosing and Treating Childhood Pneumonia and Other Childhood Illnesses inLow-Resource Settings. PloS one. 2015;10(10):e0139625.

[29] Emdin CA, Mir F, Sultana S, Kazi A, Zaidi AMK, Dimitris MC, etal. Utility and feasibility of integrating pulse oximetry into the routineassessment of young infants at primary care clinics in Karachi, Pakistan: across-sectional study. BMC Pediatrics. 2015;15(1):1-11.

[30] . Naydenova E, Tsanas A, Casals-Pascual C, Vos MD, editors. Smartdiagnostic algorithms for automated detection of childhood pneumonia inresource-constrained settings. Global Humanitarian Technology Conference(GHTC), 2015 IEEE; 2015 8-11 Oct. 2015. Abeyratne UR, Swarnkar V, Setyati A, Triasih R. Cough Sound AnalysisCan Rapidly Diagnose Childhood Pneumonia. Ann Biomed Eng. 2013;41(11):2448-62. Mulholland EK, Simoes EA, Costales MO, McGrath EJ, Manalac EM, GoveS. Standardized diagnosis of pneumonia in developing countries. PediatrInfect Dis J. 1992;11(2):77-81. Epub 1992 / 02 / 01. Thaver D, Ali SA, Zaidi AK. Antimicrobial resistance among neonatalpathogens in developing countries. Pediatr Infect Dis J. 2009;28(1 Suppl). Källander K, Nsungwa-Sabiiti J, Peterson S. Symptom overlap formalaria and pneumonia—policy implications for home management strategies.Acta Tropica. 2004;90(2):211-4. WHO. Acute respiratory infections in children: Case management insmall hospitals in developing countries: A manual for doctors and othersenior health workers. Geneva: World Health Organization; 1990. Puumalainen T, Quiambao B, Abucejo-Ladesma E, Lupisan S, Heiskanen-Kosma T, Ruutu P, et al. Clinical case review: A method to improveidentification of true clinical and radiographic pneumonia in childrenmeeting the World Health Organization definition for pneumonia. BMC InfectDis. 2008;8(1):95. {nn} group: nn1. The Acute Respiratory Infection Diagnostic Aid (ARIDA) Project,Unicef. http: / / www.unicef.org / innovation / innovation_81722.html (accessed on 08 / 03 / 2016). nn2. World's first pneumonia innovations summit unveils nextgeneration prevention, diagnostic and treatment innovations, World PneumoniaDay-2015, http: / / www.malariaconsortium.org / news-centre / worlds-first-pneumonia- innovations-summit-unveils-next-generation-prevention-diagnostic-and- treatment-innovations.htm (accessed on 08 / 03 / 2016). nn3. Pneumonia Diagnostics Project, Malaria Consortium, http: / / www.malariaconsortium.org / projects / pneumonia-diagnostics (accessed on 08 / 03 / 2016). -----------End of Appendix B------------ According to regulations, the invention has been described using language that is more or less specific to structural or methodological features. The term "comprises" and its variations, such as "comprising" and "comprised of," are always used in an inclusive sense, rather than excluding any additional features. It should be understood that the invention is not limited to the specific features shown or described, as the methods described herein include preferred forms that enable the invention to take effect. Therefore, the invention includes any forms or modifications claimed within the appropriate scope of the appended claims as properly interpreted by those skilled in the art.

[0192] Throughout the specification and claims (if any), unless the context otherwise requires, the terms “substantially” or “about” shall be understood to be not limited to the values ​​defined by the terms.

[0193] Any embodiments of the present invention are merely illustrative and are not intended to limit the invention. Therefore, it should be understood that various other changes and modifications can be made to any of the described embodiments without departing from the spirit and scope of the invention.

Claims

1. A method for automatically providing a patient's caregiver with a diagnosis of the patient's disease state, comprising the following steps: Provide diagnostic application software products that include multiple diagnostic models derived from surveys of populations including subjects with positive and non-disease states; The operation includes the input / output interface of the electronic device's memory for storing software products, prompting caregivers to identify multiple disease diagnostic parameters and values ​​for the patient; The processor of the operating electronic device selects a diagnostic model from memory based on the identified disease diagnostic parameters; The processor is operated to apply the values ​​of disease diagnostic parameters to the selected diagnostic model; and Based on the results of applying parameters to the selected diagnostic model, the diagnosis is presented to caregivers on the input / output interface so that caregivers can use it to provide treatment to patients.

2. The method according to claim 1, wherein the diagnostic parameters include: Respiratory rate, presence of fever, presence of runny nose, number of days of runny nose, number of days of cough, presence of chest retraction, temperature, BMI (body mass index), and oxygen saturation level.

3. The method of claim 1 or 2, further comprising selecting a diagnostic model by referring to one or more lookup tables, the lookup tables associating the diagnostic performance of the model with available diagnostic parameters.

4. The method according to any one of the preceding claims, comprising prompting a user to select a diagnostic model optimized for "sensitivity", "specificity" or "accuracy", wherein the method comprises operating the electronic device to select one of the diagnostic models while taking into account the optimized selection.

5. The method according to any one of the preceding claims, comprising operating the device to determine whether the value of the diagnostic parameter indicates a dangerous sign in the patient.

6. The method of claim 5, further comprising checking whether the patient's diagnostic values ​​indicate that the patient presents with general risk signs according to World Health Guidelines.

7. The method according to any one of the preceding claims includes saving the diagnostic results to a remote server, thereby enabling the diagnostic results to be saved and compared with multiple diagnostic devices.

8. The method according to any one of the preceding claims, comprising prompting the recording of at least one cough sound from the patient.

9. The method according to any one of the preceding claims, comprising requiring the recording of no more than two coughs from the patient.

10. The method of claim 8, further comprising applying at least one cough sound of the patient to a cough feature extraction engine of the diagnostic application to generate cough features.

11. The method of claim 7 or 8, wherein the patient's cough characteristics are applied to a diagnostic model to aid in diagnosis.

12. The method of claim 1, wherein the diagnostic parameters include respiratory rate, temperature, heart rate, and cough sound analysis.

13. A diagnostic device configured to prompt a clinician to input a patient's diagnostic information, and further configured to automatically present a diagnosis to the clinician, the diagnostic device comprising: An electronic memory for storing instructions includes a diagnostic application software product, the diagnostic application software product including multiple diagnostic models obtained from surveys of pneumonia-positive and non-pneumonia subjects; An electronic processor that communicates with electronic memory and executes instructions; A user interface that responds to the electronic processor for processor prompts and receiving diagnostic information; The processor is configured by the diagnostic application in use, and the user interface presents the patient's diagnosis by applying diagnostic information to at least one of the diagnostic models.

14. The method according to claim 13, wherein, The diagnostic device includes a microphone and an audio interface for connecting the microphone to an electronic processor.

15. The method of claim 13, wherein the diagnostic application includes a cough feature extraction engine, and thus in use, the processor applies the cough sound to the cough feature extraction engine to generate its cough features.

16. The method of claim 14, wherein the diagnostic device is configured by the diagnostic application in use to apply the cough feature to at least one of the diagnostic models.

17. A method for automatically diagnosing pneumonia in a patient, comprising: Use an input / output interface device to obtain the values ​​of two or more diagnostic parameters of the patient from the patient's caregiver; The diagnostic parameters include: respiratory rate, presence of fever, presence of runny nose, number of days of runny nose, number of days of cough, presence of chest retraction, body temperature, BMI (body mass index), and oxygen saturation level. Using a processor device operably connected to an input / output interface, two or more diagnostic parameters are applied to an electronic memory storing multiple pre-compiled pneumonia diagnostic models, thereby identifying the best diagnostic model among the multiple models; Apply the values ​​of two or more diagnostic signs to the identified best diagnostic model to generate diagnostic output from the diagnostic model; and The diagnostic output is operated via an input / output interface device to indicate to caregivers whether a patient has or does not have pneumonia, which caregivers can use to provide care to the patient. The pneumonia diagnostic model was derived from studies on pneumonia-positive patients and non-pneumonia-positive subjects.

Images