Peripheral neuropathy assessment method and device

Abstract

The present invention relates to a method for determining whether a person has a peripheral neuropathy. In order to be easier to generate a person using data for determination, the invention provides that sensor data from plantar pressure sensors (11, 12) used by the person is used to determine whether the person has a peripheral neuropathy.

Description

Technical Field

[0001] The present invention generally relates to the assessment of peripheral nerve function in a person. In particular, the present invention relates to a computer-implemented method, a data processing apparatus, a computer program and a system for indicating or determining whether a person suffers from peripheral neuropathy. Background Art

[0002] Methods for using video games to assess brain injuries, biomechanical problems, and other neurological disorders of the central nervous system are known from US 2012 / 0108909 A1 and US 2020 / 0129109 A1.

[0003] Nerve damage such as peripheral neuropathy can lead to serious health problems, which may lead to infection and may even require amputation, especially when the lower limbs are concerned. However, four out of five amputations can be prevented with adequate care and early diagnosis of nerve damage such as peripheral neuropathy, especially in the lower limbs.

[0004] Impaired nerve function, especially peripheral neuropathy, can be caused by diabetes. According to the International Diabetes Federation, there are currently more than 435 million people diagnosed with diabetes worldwide. This number is expected to increase to 693 million by 2045. However, even in the prediabetes stage, diabetes can lead to impaired nerve function, especially peripheral neuropathy, which often cannot be diagnosed before diabetes symptoms appear. However, by the time diabetes symptoms appear, impaired nerve function may have already led to further health problems.

[0005] Symptoms of peripheral neuropathy may usually begin distally (e.g., at the tips of the toes) and spread proximally in a symmetrical distribution. Peripheral nerves contain Aδ- and C-fibers (small, spinothalamic; temperature sensation, nociception, usually assessed by pinprick) as well as Aβ-fibers (large, dorsal nerve system; usually assessed by vibration and monofilament). In addition, reflex status may be altered (impaired Achilles tendon reflex).

[0006] Known methods of indicating or determining whether a person has peripheral neuropathy require testing by medical professionals and expensive tools or laboratory tests, or are not sensitive enough overall or to the different sizes of peripheral nerve fibers whose function may be impaired. Summary of the invention

[0007] In view of the above shortcomings, an object of the present invention may be to provide a way to indicate or determine whether a person suffers from peripheral neuropathy in an accurate, pleasant, time-saving and inexpensive manner, which allows the patient to be examined without the need for a medical professional to perform the examination and without the use of expensive equipment or laboratory tests.

[0008] According to the present invention, this purpose is achieved by the above-mentioned computer-implemented method for indicating or determining whether a person suffers from peripheral neuropathy, including loading sensor data from plantar pressure sensors, which are pressed between the person's feet and the ground, the person is playing a predetermined computer game, and at the same time, the person controls the computer game by applying pressure to the pressure sensors through the person's feet, and analyzing the sensor data using an algorithm, which is suitable for detecting whether the person suffers from peripheral neuropathy based on at least the sensor data.

[0009] For the above-mentioned data processing device having a storage device and a processor, the purpose can be achieved by: the storage device contains instructions, which, when executed by the processor, cause the processor to execute the computer-implemented method according to the present invention for indicating or determining whether a person suffers from peripheral neuropathy.

[0010] For the above-mentioned computer program comprising instructions, the object can be achieved as follows: when the program is executed by a computer, the instructions cause the computer to execute the computer-implemented method according to the present invention for indicating or determining whether a person suffers from peripheral neuropathy.

[0011] For the above-mentioned system, the purpose can be achieved as follows: the system includes a game device for playing computer games, at least two plantar pressure sensors that can be connected to the game device in a control signal transmission manner, and a data processing device according to the present invention, wherein the data processing device is suitable for loading sensor data from the plantar pressure sensors, the game device or other storage devices of the system.

[0012] By indicating or determining whether a person suffers from peripheral neuropathy based on data from a foot pressure sensor used to control a game, a person can easily and even remotely generate examination data that can indicate or determine possible peripheral neuropathy. The person does not need to be supervised while playing the game. The pressure sensor itself and especially compared to known examination devices are low cost.

[0013] The solution according to the present invention can be further improved by the embodiments mentioned below, and unless explicitly mentioned to the contrary, these embodiments can be combined as needed. These embodiments and their possible advantages are described in detail below:

[0014] According to an embodiment, the sensor data is combined and analyzed by an algorithm with game data regarding the person's game performance.

[0015] An advantage of this embodiment may be that impairment of peripheral nerve function can be indicated or determined with higher accuracy (ie, higher sensitivity and / or higher specificity).

[0016] According to an embodiment, the sensor data includes data representing the amount of pressure applied by the person and / or the timing of the pressure applied by the person, optionally combined with the timing of game control achieved by the person, which can be compared with at least one preset threshold of pressure amount and / or pressure timing.

[0017] An advantage of this embodiment may be that using data representing different aspects of the pressure exerted by a person may further improve the accuracy, ie the sensitivity and / or specificity.

[0018] According to an embodiment, the game data includes the timing and / or accuracy of positioning of an object controlling the game relative to another object of the game.

[0019] An advantage of this embodiment may be that using data representing different aspects of a person's gaming performance may further improve accuracy, ie sensitivity and / or specificity.

[0020] For example, data from sensor data and / or game data may be reaction time, anticipation time, time / frequency in / outside a predetermined optimal zone, deviation from a predetermined ideal pressure / position, normalized pressure difference, normalized pressure gradient, normalized pressure-time integral, overall feel, proficiency, muscle strength of the lower extremities, plantar pressure deviation, and / or endurance.

[0021] In particular, if sensor data is combined with game data according to an embodiment, the accuracy of the indication or determination can be improved.

[0022] According to an embodiment, the algorithm is adapted to derive a person's capabilities from sensor data and optionally game data, wherein the algorithm may be adapted to indicate or determine a person's reaction time, proficiency, muscle strength (lower limbs), plantar pressure deviation and / or endurance as their capabilities from the sensor data or a combination of sensor data and game data.

[0023] An advantage of this embodiment may be that not only derived raw data but also derived capabilities of a person controlling the game via sensors may be utilized, which may further improve the accuracy of the indication or determination.

[0024] According to an embodiment, the algorithm is an artificial intelligence model that is trained to indicate or determine whether a person suffers from impaired peripheral nerve function, particularly peripheral neuropathy, based at least on pressure data.

[0025] An advantage of this embodiment may be that during the training of the artificial intelligence model, parameters that are not considered to be indicative of peripheral neuropathy may be discovered and used by the trained artificial intelligence model. Alternatively or additionally, using the trained artificial intelligence model may further improve the accuracy of the indication or determination.

[0026] According to an embodiment, the computer game is a skill-based game.

[0027] An advantage of this embodiment may be that the skill-based game not only provides high motivation to play the game, but also provides further support by prompting the person playing the game to perform large and small movements at high and low intensities, so that different fibers and reflexes are applied when playing the skill-based game, which helps to indicate or determine the impaired function of fibers of different sizes.

[0028] According to an embodiment, the system includes an insole or a shoe, wherein the insole of the shoe insole respectively includes at least one plantar pressure sensor. Alternatively, the plantar pressure sensor can be fixed to the sole of the person's foot, wherein this embodiment may require a third party to assist in the fixing of the sensor. As a further alternative, the plantar pressure sensor can be set in a sock or at a location where the sock contacts the sole of the person's foot when the subject wears the sock.

[0029] For example, a pair of shoes equipped with sensor-equipped insoles provided by AntiSense System of IEE SALuxembourg can be used. These insoles contain an electronic control unit (ECU) and eight pressure sensors (5.57 cm 2 High dynamic HD002 force sensing resistor), which can be integrated into the heel, lateral arch, 1st, 3rd and 5th metatarsal, big toe and toe area of ​​the foot. The sensor allows to detect pressure in the range of 250mbar to 7bar with a sensitivity of 3.4mbar. The ECU can include a nine-degree-of-freedom inertial measurement unit (IMU; embedded with 3-axis accelerometer, 3-axis gyroscope and 3-axis magnetometer), sampling rate up to 500Hz, data synchronization (between insole and smart device), automatic identification of the foot side, 16GB internal storage and energy supply up to 10 hours. The sensor can be embedded in the film and will not protrude. The insole is integrated with the footwear using a support layer and a sponge protective layer of ethylene-vinyl acetate-30 material.

[0030] The sensor data may be recorded at over 50 Hz, over 100 Hz, up to or over 200 Hz, and transmitted in real time to a gaming device, such as via Bluetooth or a cable, for smooth game control. A person utilizing the system for indicating or determining whether a person suffers from peripheral neuropathy may sit in an armless chair in front of a table with a gaming device, such as a tablet, (e.g., Samsung Galaxy Tab A T580 or iPad) placed on the table, connected to the insole via Bluetooth 5.0 or a cable. The person may initially calibrate the insole through eight predetermined steps.

[0031] An advantage of this embodiment may be that no dedicated hardware needs to be developed, but existing hardware may be used, which further reduces costs and improves availability.

[0032] For example, the pressure threshold measured during the calibration of the plantar pressure sensor can be used to normalize the absolute pressure value to a range of 0 to 1. Subsequently, the person can be familiar with the setting and control techniques of the video game through a standardized tutorial repeated as needed. Each game session may consist of at least one or all of the following four games: Apple Game (AC), Balloon Flight Game (BF), Cross Pressure Game (CP) and Island Jumping Game (IJ), which can be played in sequence. Each session can last for at least 5 minutes, up to 10 minutes, up to 15 minutes or up to 20 minutes. The sensor data and game results obtained can be transmitted to a remote server in a complete data packet for data visualization and analysis, or can be transmitted to another computer for data visualization and analysis, or can be analyzed and / or visualized by the game device playing the game. The spider graph can provide feedback on the game performance of the person after the session (exemplary key abilities: reaction time, feeling, proficiency, muscle strength (lower limbs), plantar pressure deviation and / or endurance).

[0033] The data processing device and / or gaming device may be a computer having a processor, solid-state non-volatile memory, RAM, a data interface (such as USB, HDMI, DisplayPort, etc.), and may in particular be any personal computer, desktop, laptop, tablet, smartphone, smartwatch or server, etc.

[0034] So, in other words, peripheral neuropathy (PNP) is a common comorbidity in patients with diabetes or metabolic syndrome (MetS) and can be diagnosed by clinical examinations (e.g. pinprick, vibration perception, Tip Therm, monofilament, reflexes). These are very time consuming and rarely performed by physicians in primary care centers, probably due to time constraints.

[0035] Early identification of PNP is a hallmark for preventing complications (e.g., foot ulcers). Clinical examination includes standardized procedures (pinprick, vibration perception, tip thermo, monofilament, reflexes). These analyses are susceptible to investigator-related bias and are performed with low frequency in primary practice, probably due to time constraints.

[0036] Diagnosis of PNP is challenging for primary care clinicians due to its diverse forms and presentations. Particularly in patients with diabetes or MetS, the onset is insidious and most affected individuals are unaware of distal symmetric PNP. Other common causes consist of excessive and chronic alcohol consumption, nerve damage, toxin exposure, malnutrition, chemotherapy, and genetic disorders.

[0037] According to the International Diabetes Federation (IDF), more than 435 million people worldwide are diagnosed with diabetes, a number that is expected to rise to 693 million by 2045. PNP affects 25% to 50% of people with diabetes, with severity varying depending on age, duration of diabetes, and level of glycemic control. PNP severely impacts patients’ mobility and is associated with gait and coordination disturbances, increased likelihood of falls, higher rates of diabetic foot syndrome, and fragile mental health conditions. In addition to impaired sensation, PNP can cause symptoms of discomfort and pain. It is estimated that 19% to 34% of people with diabetes are affected by complications such as tissue damage, infection, and ultimately minor and major foot ulcers during their lifetime, especially when PNP diagnosis is delayed or preventive measures are inadequate. Notably, one in five moderate to severe diabetes-related infections leads to lower limb amputation. On the other hand, four out of five amputations could be prevented with proper foot care.

[0038] All these aspects call for optimization of PNP diagnosis and possibly even prevention and management. Due to the nature of the disease, prompt diagnosis and repeated monitoring of high-risk individuals are mandatory, considering the insidious onset and diverse clinical presentation of PNP. Up to 50% of affected individuals remain asymptomatic, while the rest experience numbness, tingling, pain, or weakness. Symptoms typically begin distally (i.e., the tips of the toes) and spread proximally in a symmetrical distribution. The peripheral nerves include large nerve fibers (Aα- and Aβ-fibers; vibration and pressure sensation) and small nerve fibers (unmyelinated C- and Aδ-fibers; pain and cold and heat sensation).

[0039] Nerve conduction studies (NCS) and skin biopsy are not suitable as screening tools due to their invasive nature. According to a position statement published 15 years ago by the American Diabetes Association (ADA), “the evaluation should include a careful history and temperature or pinprick sensation [..] and vibration sensation [..]”. Some innovations have been proposed as point-of-care devices, however, none have become routine. As PNP diagnostics are seen as a game-changer for intervention and prevention of complications, innovative strategies are greatly needed.

[0040] According to the present invention, a game-based examination platform is provided, which consists of four video games, guided by a shoe insole equipped with a pressure sensor as a control unit. In patients diagnosed with diabetes or MetS, game performance is correlated with clinical PNP findings through neuropathy symptom scores, neuropathy degree scores, and NCS.

[0041] A fun neurological function assessment device was established and evaluated through a video game controlled by foot motion and plantar pressure sensors (e.g., insoles equipped with pressure sensors). Pressure data can be transmitted to a central database in real time for analysis. The game can be designed with different challenges, and the feature extraction method can be applied and validated in a pilot exploratory study. The game-based method was evaluated for correlation with NCS and clinical findings (NDS, NSS) through predefined hypothesis-driven parameters (reaction time, sensation, proficiency, muscle strength (lower limbs), plantar pressure deviation, and endurance). Exemplary representative game features indicating significant differences between patients with small-fiber or large-fiber peripheral neuropathy (SFN, LFN) were identified. In addition, an exemplary training and optimization classification model and clinical phenotype of PNP were reported using selected game features.

[0042] The game-based approach identified patients with PNP based on their gaming performance. Patients were tasked with playing a series of digital games guided by insoles equipped with pressure sensors. When correlated with clinical neuropathy scores and nerve conduction studies, the developed algorithm played the game with 76.1% and 81.7% accuracy for the left and right feet, respectively. In addition, certain features of the gaming sessions enabled fine-tuning of the PNP impairment pattern (small, large, or mixed nerve fiber dysfunction).

[0043] For example, of the 329 individuals, 247 (75.1%) were diagnosed with clinically definite PNP. Asymmetric findings were detected in 88 (26.7%) individuals. In the subgroup using NCS as the gold standard, gaming performance was significantly correlated with sensory and motor nerve conduction velocity (NCV) and the amplitude of major nerves in the lower extremities (R = 0.65, p < 0.001, adjusted R2 = 0.36). In the age-matched subgroup 2, it was determined that PNP severity was significantly associated with hypothesis-driven key abilities of gaming performance. The learning model derived from the extracted gaming features achieved an accuracy of 76.1% (left foot) and 81.7% (right foot) for detecting PNP. When evaluating the PNP injury pattern (small, large or mixed nerve fiber injury), the resulting multi-classification model yielded a multi-classification AUC of 0.76 (left foot) and 0.72 (right foot).

[0044] Therefore, the game-based neuropathy assessment platform provides meaningful assessment of PNP and is eligible for examination-independent testing in primary care.

[0045] The inventors have found that the design of a standardized video-based game can yield important information about the patient's neuropathy status. The overall complex interaction of game perception and execution is achieved in a standardized environment through plantar pressure sensors, such as insoles equipped with pressure sensors, which generate data sets from sensor data for sample analysis. The selected method allows access to these data for the classification of patients with different degrees of PNP and small / large nerve fiber impact into different subgroups. Even as a telemedicine application, it is feasible to screen patients for PNP in an entertaining way. Therefore, in an experimental setting, remoteness is feasible and provides a simplified tool that is easy to adopt in primary care.

[0046] To develop an exemplary suitable artificial intelligence (AI) algorithm, game parameters extracted from the obtained dataset were used to derive a binary classification model to identify feet with PNP in individuals diagnosed with diabetes or metabolic syndrome.

[0047] For example, for AI modeling, the obtained dataset was initially divided into a test dataset and a training dataset (ratio 3:7). Random sampling was performed in each classification, and the overall categorical distribution of the data was retained. Only the training dataset was used to identify features with significant inter-group differences and imported into the classifier for evaluating the importance of features based on model-independent metrics. Subsequently, models containing different subsets of the top-ranked features (i.e., differences in the number and / or order of features) were tested. During the modeling process, multiple classifiers were tested, including random forest (RF), lasso and elastic net regularized generalized linear model (GLMNET), support vector machine (SVM), and gradient boosting (GBM). The area under the receiver operating characteristic curve (AUC-ROC) was selected as a performance metric to compare models trained with different feature combinations. Five-fold ten-fold repeated cross validation was used in training to avoid overfitting and obtain more accurate estimates of model performance. This statistical method repeatedly divides the training dataset into five subsets of approximately equal size three times. Each subset contains the same proportion of labels as the full dataset. Four of the five subsets are used for model training, and the remaining subset is used for validation. The mean AUC of the cross validation was considered in the grid search of the parameter combination that best improved the model performance. Finally, the obtained model was further applied to the hold-out test dataset to evaluate the predictive performance of the model. The classification performance was evaluated by calculating the AUC-ROC, balanced accuracy, sensitivity, and specificity, as compared with the reference standard of clinical examination.

[0048] In addition, multiple classification models were trained to predict small, large, or mixed nerve fiber damage in the foot of patients with diabetes or metabolic syndrome. All datasets were used for modeling, and due to limited sample size and unbalanced classification distribution, a five-fold ten-fold repeated cross validation was used. Multi-classification AUC was calculated to evaluate model performance. R programming language (version 4.2.1) and related open source libraries were used for statistical computing and machine learning algorithms.

[0049] During the development phase, the game features were analyzed as follows: NCS was performed on 37 patients (subgroup 1), all of whom were diagnosed with diabetes and had normal cognitive status (MoCA ≥ 26). The distribution of age, weight, BMI, and neuropathy score in this subgroup was similar to the entire group. In subgroup 1, the gender distribution was slightly biased towards female patients. Patients' game performance was significantly correlated with sensory and motor conduction velocity and amplitude in the lower limbs. Feature extraction of different parameters from four games was set for each foot. Overall, 277 independent game features showed significant associations with NCS parameters. A correlation between the motor nerve conduction amplitude and the maximum execution time of the right foot was observed in task 8 of the island jumping game (R = 0.65, p < 0.001, adjusted R2 = 0.36). The BF and AC games provided more features related to the NCV and amplitude of the left foot. However, more features extracted from the BF and IJ games were related to the NCV and amplitude of the right foot. 164 and 113 independent game features were significantly associated with motor and sensory nerve status, respectively. We found 146 independent gaming features were associated with NCV and 131 were associated with foot nerve amplitude.

[0050] Of all individuals (n=329), a subgroup 2 of patients aged >55 years and with normal cognitive status (MoCA≥26) was generated. No significant differences in age, sex, and BMI were observed between patients with and without clinical PNP. Individuals with peripheral neuropathy had a significantly longer duration of diabetes and more severe neuropathy symptoms. Clinically, evidence of PNP was present in 118 left feet and 120 right feet. For all of these feet, at least one site of insensitivity was found during the 10 g monofilament test. PNP was found in both feet of 112 patients. No PNP was present in 47 patients, and unilateral PNP was present in 26% of patients.

[0051] Successful game performance requires key skills related to neurological function. For example, four hypothesis-driven key abilities were defined: reaction time (comprehension; immediate response to the task), sensation (fine-tuning of pressure application within subtasks), proficiency (overall completion of each game), lower extremity muscle strength (completion with high pressure application), plantar pressure deviation (pressure distribution between left and right feet), and endurance (constancy of pressure application within the task). All key abilities showed significant differences between patients without PNP (n = 47) and those with symmetric PNP (n = 112).

[0052] Using the game features extracted from 173 patients in subgroup 2, the prediction model of PNP was trained using five-fold ten-fold repeated cross validation. For example, 70% of the game data set was used for model training and 30% for testing. In the reserved test data set, the obtained lasso and elastic network regularized generalized linear model (GLMNET) identified left foot PNP with an adjusted accuracy of 76.1% (sensitivity 77.1%, specificity 75.0%, AUC-ROC 0.72). For the determination of right foot PNP, another GLMNET model achieved an accuracy of 81.7% (sensitivity 83.3%, specificity 80.0%, AUC-ROC 0.75).

[0053] In addition, for example, multi-classification models were built to distinguish between feet without PNP (w / o) and SFN, LFN, and mixed nerve fiber damage. Exemplary game features for the left and right feet were extracted from the game dataset. The trained gradient boosting model (GBM) obtained a multi-classification AUC of 0.76 and 0.72 when identifying the nerve fiber damage pattern of the left and right feet, respectively. These models performed better in distinguishing SFN and LFN.

[0054] Therefore, in summary, game-based settings to detect PNP in patients with diabetes or MetS are feasible and accurate. Games can be designed to detect key abilities related to neurological function, and the interpretation of results is based on algorithms developed based on hypothesis-driven assumptions and AI methodology. Findings from NCS data suggest that standardized game sessions lasting 5 to 25 minutes (e.g., 15 minutes) are sufficient to obtain an objective assessment of peripheral nerve function. The extracted game features were associated with the amplitude of NCV and the main nerves of the foot, indicating a possible combination of axonal degeneration and segmental demyelination. The learning model trained with a game dataset from an age-matched subpopulation was able to distinguish between feet with PNP and healthy feet, with an accuracy of 76.1% and 81.7% for each foot, respectively. The definition of PNP is based on widely used diagnostic criteria (a combination of NSS / NDS), but is implemented separately for each foot. 75.1% of subjects were determined to have clinically clear PNP, indicating that the risk of developing foot complications in this group is extremely high. The selected diagnostic criteria based on clinical examination are able to distinguish the nerve damage pattern of PNP. The obtained multiclassification model predicted the nerve injury pattern (small, large, or mixed fiber neuropathy) with an AUC of 0.76 and 0.72 for the left and right feet, respectively.

[0055] Asymptomatic neuropathy was observed in 30% of the subjects. This emphasizes the importance of developing such an easy-to-use screening tool to avoid or delay the progression of PNP. In addition, more than a quarter of the subjects were diagnosed with asymmetric neuropathy. To avoid methodological errors caused by combining the analysis with asymmetric neuropathy, each foot can be analyzed and modeled separately.

[0056] The game-based approach can be performed with a high degree of standardization, largely independent of the researcher, while requiring no healthcare professional. The study had a low threshold for completion, even for those with little or no gaming experience. The immediate response of the subjects after completing the game was very positive. More than 90% wanted to repeat the session. However, the results of the second gate were not included in the study results disclosed in this paper. As a visionary prospect in the context of telemedicine, individuals can perform the gaming session at home, with the distribution of the results after the AI ​​model prediction being distributed remotely to the attending physician. This approach may be applicable to high-risk populations of older adults with limited mobility or geographical restrictions in primary care settings.

[0057] In this application, the following abbreviations may be used:

[0058] AC: Take the Apple

[0059] ADA: American Diabetes Association

[0060] AUC-ROC: Area under the receiver operating characteristic curve

[0061] BF: Balloon Flight

[0062] CP: Cross Pressure

[0063] PNP: Peripheral Neuropathy

[0064] ECU: Electronic Control Unit

[0065] GBM: Gradient Boosting Model

[0066] GLMNET: Lasso and Elastic Net Regularized Generalized Linear Models

[0067] IDF: International Diabetes Federation

[0068] IMU: Inertial Measurement Unit

[0069] IJ: Island Hopping

[0070] LFN: Large fiber neuropathy

[0071] NDS: Neuropathy Severity Score

[0072] NSS: Neuropathy Symptom Score

[0073] NYHA: New York Heart Association

[0074] NCS: Nerve Conduction Study

[0075] PNP: Peripheral Neuropathy

[0076] SD: Standard deviation

[0077] SFN: Small fiber neuropathy

[0078] SVM: Support Vector Machine

[0079] TCL: Left Foot Mission Combination

[0080] TCR: Mission Combination for the Right Foot BRIEF DESCRIPTION OF THE DRAWINGS

[0081] The present invention is described in more detail and exemplarily below by means of advantageous embodiments and with reference to the accompanying drawings. The embodiments are only possible configurations, wherein the individual features described above may be provided separately or may be omitted in the drawings:

[0082] For elements of exemplary embodiments that correspond in form and / or function to elements of other exemplary embodiments, the same reference numerals are used.

[0083] Figure 1 An exemplary embodiment of a computer-implemented method for indicating or determining whether a person suffers from peripheral neuropathy is shown, having a step including collecting sensor data, which is not part of the claimed invention.

[0084] Figure 2 An exemplary embodiment of a system for indicating or determining whether a person suffers from peripheral neuropathy is shown.

[0085] Figures 3 to 7 An exemplary embodiment of a game played by a person using a foot pressure sensor is shown.

[0086] Figure 8 An exemplary embodiment of a classification modeling scheme for artificial intelligence is shown.

[0087] Fig. 9 The functionality of an exemplary embodiment of a computer method implemented with artificial intelligence is schematically illustrated.

[0088] Fig.10 An overview of an exemplary embodiment of a system for indicating or determining whether a person suffers from peripheral neuropathy is schematically illustrated.

[0089] Fig.11 The performance of nerve conduction studies and correlation with game-based findings are schematically shown.

[0090] Fig.12 Schematic illustration of possible classifications of peripheral neuropathy based on game-generated sensor data.

[0091] Fig.13 Another possible classification of peripheral neuropathy based on game-generated sensor data is schematically shown.

[0092] Fig.14 The calibration procedure of the plantar pressure sensor is schematically shown.

[0093] Fig.15 Schematic illustration of feature extraction and analysis from game-generated sensor data for use in a peripheral neuropathy classification model.

[0094] Fig.16 Consideration of key functions, game performance, and scoring of sensor data generated by the game is schematically shown. DETAILED DESCRIPTION

[0095] According to the present invention and as a flow chart, Figure 1 A first exemplary embodiment of a computer-implemented method for indicating or determining whether a person suffers from peripheral neuropathy is shown.

[0096] The computer-implemented method 1 for indicating or determining whether a person suffers from peripheral neuropathy begins with a first method step 1. Method steps 3 and 4 involve contacting at least one plantar pressure sensor with each foot of the person (step 3), and having the person play a game while collecting pressure and optional game data (step 4). Method steps 3 and 4 are not part of the claimed invention, but are useful for understanding the claimed invention. Steps 3 and 4 can be supervised by another person, who can even assist the person who may suffer from peripheral neuropathy, wherein the other person need not be any medical professional.

[0097] In method step 5 , the sensor data and optionally the game data are analyzed by an algorithm, which is suitable for identifying whether a person suffers from peripheral neuropathy based on the sensor data.

[0098] Finally, method 1 may end with method step 6, in which the indication or determination is communicated to the person or other person, such as a medical professional who normally examines or treats the person, such as a general practitioner or specialist. The other person need not be present to apply the sensor or to be present when playing the game.

[0099] Figure 2 An exemplary embodiment of a system for indicating or determining whether a person suffers from peripheral neuropathy according to the present invention is schematically illustrated.

[0100] The system 10 includes at least two plantar pressure sensors 11, 12, each of which is applied to the sole of a person's foot so that the person can control the game by pressing the foot on the ground to varying degrees. The system 10 also includes a game device 13 for playing computer games. The plantar pressure sensors 11, 12 can be connected to the game device 13 in a control signal transmission manner. The control signal depends on the pressure applied to the ground by the plantar pressure sensors 11, 12, and can be proportional to the pressure, for example. The control signal can be converted into sensor data by the plantar pressure sensors 11, 12 or the game device 13. Optionally, the plantar pressure sensors 11, 12 can be connected to a data processing device 14 of the system 10 in a signal or, for example, data transmission mode. Therefore, the data processing device 14 can be suitable for loading sensor data from the plantar pressure sensors 11, 12, the game device 13 or other storage devices of the system 10. The data processing device 14 may be the gaming device 13 or a part thereof, may be arranged separately from the gaming device 13, and may be directly connected to the gaming device 13 and / or the plantar pressure sensors 11, 12, for example wirelessly or via a cable or via the Internet.

[0101] Figure 3 An exemplary embodiment of a plantar pressure sensor 11 , 12 is shown schematically.

[0102] The plantar pressure sensors 11, 12 may each include at least one or more pressure-sensitive elements 20, i.e., pressure sensor elements. The plantar pressure sensors 11, 12 may be insoles. The plantar pressure sensors 11, 12 may include a common signal-to-data converter 21, which is connected to the pressure-sensitive element 20 and converts the sensor signal into sensor data. The signal-to-data converter 21 may be an analog-to-digital converter. Figure 3 The exemplary embodiment of FIG. 4 shows two plantar pressure sensors 11, 12, each comprising a signal to data converter 21. The plantar pressure sensors 11, 12 may comprise a signal or data interface 22 adapted to output a sensor signal or sensor data. The signal interface or data interface 22 may be a physical interface, such as a connector. Figure 3 The exemplary embodiment of FIG. 4 shows two plantar pressure sensors 11 , 12 with a wireless signal or data interface 22 , which can output the sensor data in a wireless manner, for example via Bluetooth.

[0103] Figures 4 to 7 Schematically illustrated are games and how they may be controlled to generate data which is used according to the invention to indicate or determine whether a person suffers from peripheral neuropathy.

[0104] The measured pressure thresholds can be used to normalize the absolute pressure values, for example, to a range of 0 to 1. Subsequently, the person can become familiar with the setup and control techniques of the computer game through a predetermined tutorial that can be repeated as needed. Each gaming session can include multiple games, for example, four: Connect the Apple (AC, Figure 4 ), Balloon Flight (BF, Figure 5 ), cross pressure (CP, Figure 6 ) and island hopping (IJ, Figure 7 ), these games can be played selectively or sequentially. Each game may last for several minutes, such as up to 5 minutes, up to 10 minutes, up to 15 minutes, or about 20 minutes. The acquired sensor data and game data can be transmitted to a remote server in a complete data package for data visualization and analysis.

[0105] In the AC game, a person controls a car to catch apples that fall from a tree. The BF game challenges a person to fly a balloon along a predetermined path. The CP game requires a person to provide a predetermined load through a predetermined part of the foot in a predetermined pattern. In the IJ game, a person controls a virtual character to jump from one island to another, wherein at least one pressure variable such as the magnitude of the pressure, the timing of the pressure, and the duration of the pressure determines whether the virtual character falls on another island. A known game with a similar approach but using a different control mechanism may be "Hot Lava" by Klei Entertainment Inc., which is controlled by a game handle or a touch screen (i.e., manually). All movements of objects (such as a car in the AC game, a balloon in the BF game, a virtual character in the IJ game, and / or an indicator on the CP game) are controlled by sensor signals, which rely on a person applying pressure on a pressure sensor through his or her foot.

[0106] Game data extraction from different game parameters of game tasks defined and game task combinations of games can be set. Game parameters can be calculated from each game task and regarded as the main feature, such as the reaction time in the first task of the AC game. In addition, the concept of task combination (TC) can be introduced, that is, a group of game tasks with similar technical parameters or macro measurements can be combined. For example, TC1 of the AC game can contain all tasks of the game. TC2 can include all tasks of using the left foot to control the movement of the car to catch the apple. The corresponding TC3 can be composed of all tasks using the right foot. The sum, mean value and standard deviation of the main features of the game tasks of TC can be regarded as secondary features. For example, in the AC game, the reaction time of TC1 can be a secondary feature, which can be calculated from the average reaction time of all tasks. In general, each data set (for each person and all four games in total) can extract 1880 unique parameters reflecting the performance of people in overall and similar tasks.

[0107] The games may provide a data set comprising sensor data and optionally game data, wherein the AC game may provide 583 data elements or features, the BF game may provide 528 data elements or features, the CJ game may provide 473 data elements or features, and the IJ game may provide 296 data elements or features, such that if a person plays all four games, there are a total of 1,880 data elements or features in the data set of sensor and optionally game data to indicate or determine whether the person suffers from peripheral neuropathy.

[0108] and Figure 5 compared to, Figure 5a Additionally shown is region D, which shows the position data of the virtual balloon controlled by the right foot compared to the optimal position. Figure 7 compared to, Figure 7a Additionally shown is region E, which shows right foot data for the island jumping game.

[0109] Figure 8 A classification scheme is shown for an exemplary embodiment of an algorithm representing an AI (artificial intelligence) model for indicating or determining whether a person suffers from peripheral neuropathy.

[0110] For example, an AI (artificial intelligence) model can be derived from the obtained dataset to distinguish between healthy controls and patients with diabetes and sensory neuropathy in an age-matched group (Group 2). NDS (Neuropathy Deficit Score) can be used to generate classification labels for supervised AI learning. NDS greater than or equal to 2 and impaired vibration sensation (0-4 / 8) can be considered as the presence of neuropathy (DPN), while NDS = 0 can be considered as the absence of neuropathy (control group). Secondly, in the DPN group, a more detailed phenotypic analysis can be achieved by analyzing the dysfunction of different fiber types. Figure 8 The details of these modeling scenarios are summarized in Figure 8 In model 2, it can be assumed that patients with reduced / lost pinprick (nociception) or temperature sensation have severe damage to the Aδ- / C-fibers. Figure 8 In model 3, severe Aβ-fiber polyneuropathy can be assumed by impaired vibration sensation (less than 3 / 8) or abnormal 10 g monofilament test results (diminished or absent). The remaining patients in the DPN group can be classified as moderate Aβ-fiber polyneuropathy (impaired vibration sensation: 3-4 / 8). Figure 8 In model 4, the absence of Achilles tendon reflex can be regarded as a positive label during model training. Normal or impaired reflexes are marked as negative labels.

[0111] For feature selection, the obtained dataset can first be divided into a test dataset and a training dataset (e.g., a ratio of 3:7). Random sampling is performed in each classification, and the overall classification distribution of the data is preserved. Only the training dataset can be imported into the classifier, and the importance of the features is estimated based on model-independent indicators (e.g., the area under the receiver operating characteristic curve). Next, models with different subsets of the top-ranked features (i.e., differences in the number and / or order of features) can be tested. If the dataset under study is unbalanced, Cohens Kappa (classification accuracy normalized by the imbalance of the classification in the data) can be selected as a performance indicator to compare models trained with different feature combinations. Models with higher Cohens Kappa values ​​can be selected as candidate models.

[0112] For model training, a support vector machine (SVM) with three kernel functions (linear, radial, and polynomial) can be selected. Clinical studies have shown that SVM models perform better in medical classification problems with limited and unbalanced data sets. Three repeated five-fold cross-validation can be used in feature sorting and training to avoid overfitting and obtain more accurate estimates of model performance. This statistical method can repeatedly divide the training data set into, for example, five subsets of approximately equal size three times. Each subset can contain the same proportion of labels as the complete data set. Subsets, for example, four of the five subsets can be used for model training, while the remaining subsets can be used for validation. In the grid search for the parameter combination that best improves model performance, the average accuracy of cross-validation can be considered. Finally, the obtained candidate model can be further applied to the test data set to evaluate the predictive ability of the model.

[0113] The R programming language (version 4.0.4) and related open source libraries can be used for statistical computing and machine learning algorithms.

[0114] For example, the following mentions possible libraries to use (package, version, description, source, release date):

[0115] Caret, 6.0-86, Miscellaneous functions for training and plotting classification and regression models, Comprehensive R Archive Network (CRAN), 2020-03-20 10:20:07 UTC

[0116] Colorspace, 2.0-0, "Maps between various color spaces including RGB, HSV, HLS, CIEXYZ, CIELUV, HCL (Polar CIELUV), CIELAB, and Polar CIELAB." Qualitative, sequential, and diverging palettes based on HCL colors and corresponding chromaticities for ggplot2 are provided. Palette selection is assisted by an interactive application (with Tcl / Tk or Shiny GUI) and a Shiny application with an HCL color picker and a color vision deficiency simulator. Plotting functions for displaying and evaluating palettes include color swatches, visualization of the HCL space, and trajectories in the HCL and / or RGB spectra. Color manipulation functions include: desaturation, brightening / darkening, blending, and simulation of color vision deficiencies (deuteranopia, protanopia, tritanopia). Details are available on the project webpage. http: / / colorspace.R-Forge.R-project.org / and the accompanying scientific paper Zeileis et al. (2020, Journal of Statistical Software,<doi:10.18637 / jss.v096.i01> ). ”, ComprehensiveRArchive Network (CRAN), 2020-11-1107:40:15UTC

[0117] Corrplot, 0.84, "Graphical display of correlation matrices or matrices in general. It also includes some algorithms for reordering matrices. In addition, corrplot excels at details, including choice of colors, text labels, color labels, layout, etc.", Comprehensive R Archive Network (CRAN), 2017-10-16 22:58:08UTC

[0118] Dfidx, 0.0-4, provides extended data frames with special data frame columns containing two indexes, possibly with nested structures, Comprehensive R Archive Network (CRAN), 2021-02-03 05:40:10UTC

[0119] Dplyr, 1.0.7, "Fast and consistent tools for working with data frame-like objects both in-memory and out-of-memory.", Comprehensive R Archive Network (CRAN), 2021-06-18 23:20:01UTC

[0120] Egg, 0.4.5, Various functions to help customize 'ggplot2' objects. Provides high-level functions to post-process 'ggplot2' layouts and allow alignment between plot panels, as well as setting panel sizes to fixed values. Other functions include custom 'geom' and helper functions to force symmetric scales or add labels to faceted plots, Comprehensive R Archive Network (CRAN), 2019-07-13 06:00:27UTC

[0121] Formula, 1.2-4, "Basic structure of extended formulas with multiple parts on the right and / or multiple responses on the left (see<doi:10.18637 / jss.v034.i01> ). ”, Comprehensive R Archive Network (CRAN), 2020-10-1613:50:06UTC

[0122] Gbm, 2.1.8, "Implements the AdaBoost algorithm of Freund and Schapire and an extension of Friedman's gradient boosting machine. Includes regression methods for least squares, absolute loss, t-distribution loss, quantile regression, logistic, multinomial logistic, Poisson, Cox proportional hazards partial likelihood, AdaBoost exponential loss, Huberized hinge loss, and a learning to rank metric (LambdaMart). Originally developed by Greg Ridgeway.", Comprehensive R Archive Network (CRAN), 2020-07-15 10:00:02 UTC

[0123] GGally, 2.1.0, "The R package 'ggplot2' is a plotting system based on the grammar of graphics. 'GGally' extends 'ggplot2' by adding several functions to reduce the complexity of combining geometric objects with transformed data. Some of the functions include pairwise plot matrices, two sets of pairwise plot matrices, parallel coordinates plots, survival analysis plots, and several functions to draw networks.", Comprehensive R Archive Network(CRAN), 2021-01-06 01:20:06UTC

[0124] Ggnewscale, 0.4.5, Using multiple fills and chromaticities in 'ggplot2', Comprehensive RArchive Network (CRAN), 2021-01-11 08:00:35UTC

[0125] ggplot2, 3.3.5, "A system for 'declarative' creation of graphics based on the "grammar of graphics". The user provides the data, tells 'ggplot2' how to map variables to aesthetics, which graphical primitives to use, and it takes care of the details.", Comprehensive RArchive Network (CRAN), 2021-06-25 11:30:09UTC

[0126] ggpubr, 0.4.0, "The 'ggplot2' package is excellent and flexible for elegant data visualization in R. However, the default generated plots require some formatting before we can send them for publication. In addition, for customizing 'ggplot', the syntax is opaque, which increases the difficulty for researchers without advanced R programming skills. 'ggpubr' provides some easy-to-use functions for creating and customizing publishable plots based on 'ggplot2'.", Comprehensive RArchive Network(CRAN), 2020-06-27 06:20:02UTC

[0127] ggrepel, 0.9.1, "Provides geometries for text and labels for 'ggplot2', helping avoid overlapping text labels. Labels are mutually exclusive and kept away from data points.", Comprehensive R Archive Network (CRAN), 2021-01-15 22:00:02UTC

[0128] glmnet, 4.1, Very efficient procedures for fitting linear regression, logistic and multinomial regression models, Poisson regression, Cox models, multi-response Gaussian, and grouped multinomial regression with lasso or elastic net regularization paths. There are two new important additions. The series argument can be a GLM Series object, which opens the door to any programmed series. This comes at a modest computational cost, so when built-in series are sufficient, they should be used instead. Another novelty is the relaxation option, which refits each set of activities in the path without penalty. The algorithm uses cyclic coordinate descent in a path-wise manner, as described in the paper listed at the following URL, Comprehensive R Archive Network (CRAN), 2021-01-11 08:00:30UTC

[0129] gmodels, 2.18.1, Various R programming tools for model fitting, Comprehensive R Archive Network (CRAN), 2018-06-25 16:23:05UTC

[0130] gridExtra, 2.3, "provides extensive user-level functionality for working with "grid" graphics, in particular arranging multiple grid-based plots on a page, and drawing tables.", Comprehensive R Archive Network (CRAN), 2017-09-09 14:12:08 UTC

[0131] Hmisc, 4.4-2, "Contains many functions for data analysis, advanced graphics, utility operations, functions for calculating sample size and power, importing and annotating datasets, imputing missing values, advanced table making, variable clustering, string manipulation, and conversion of R objects to LaTeX and HTML code.", Comprehensive R Archive Network (CRAN), 2020-11-29 06:00:16UTC

[0132] Imager, 0.42.3, "Fast image processing for images in up to four dimensions (two spatial dimensions, a time / depth dimension, and a color dimension). Provides traditional image processing tools (filtering, morphology, transformations, etc.) as well as a variety of functions that facilitate the analysis of image data using R. Packages include 'Clmg',<http: / / cimg.eu> , a simple, modern C++ library for image processing. ”, Comprehensive R Archive Network (CRAN), 2020-05-11 16:40:06UTC

[0133] Jpeg, 0.1-8.1, This package provides convenient methods to read, write, and display bitmap images stored in JPEG format. It can read and write raw vectors in files and memory. , Comprehensive R Archive Network (CRAN), 2019-10-24 14:51:52UTC

[0134] Knitr, 1.31, "A general-purpose tool for dynamic report generation in R using literate programming techniques.", Comprehensive R Archive Network (CRAN), 2021-01-27 15:10:05UTC

[0135] Lattice, 0.20-41, "A powerful and elegant high-level data visualization system inspired by Trellis graphics, focused on multivariate data. Lattice is powerful enough for typical graphical needs, yet flexible enough to handle most nonstandard requirements. See ? Lattice for an introduction.", Comprehensive R Archive Network (CRAN), 2020-04-02 12:00:06 UTC

[0136] Libcoin, 1.0-8, "Infrastructure for linear test statistics and permutational inference in the framework of Strasser and Weber (1999)<https: / / epub.wu.ac.at / 102 / > . This package is not intended for use by end users. The ComprehensiveR Archive Network (CRAN) package 'coin' implements all user interfaces and is available to anyone. ”, Comprehensive R Archive Network (CRAN), 2021-02-08 11:00:03UTC

[0137] Magrittr, 2.0.1, "provides a mechanism for chaining commands via a new forward pipe operator %>%. This operator passes a value or the result of an expression into the next function call / expression. Flexible support for the type of the right-hand side expression is available. See the package documentation for more information. To quote Rene Magritte: Ceci n'est pas un pipe.", Comprehensive R Archive Network (CRAN) 2020-11-17 16:20:06UTC

[0138] Matrix, 1.3-2, "A rich hierarchy of matrix classes, including triangular, symmetric, and diagonal matrices, dense and sparse and with mode, logical, and numeric entries. Utilizes the 'LAPACK' and 'SuiteSparse' libraries for a variety of methods and operations on these matrices.", Comprehensive R Archive Network (CRAN), 2021-01-06 17:40:19UTC Mboost, 2.9-4, "The functional gradient descent algorithm (boosting) is used to optimize a general risk function, using component-wise (penalized) least squares estimators or regression trees as base learners for fitting generalized linear, additive, and interaction models to high-dimensional data. The models and algorithms are described in \doi{10.1214 / 07-STS242}, and a practical tutorial is available from \doi{10.1007 / s00180-012-0382-5}. The package allows the user to specify the loss function and base learners.", Comprehensive R Archive Network (CRAN), 2020-12-10 09:40:02UTC

[0139] MLeval, 0.3, Direct and detailed evaluation of machine learning models. 'MLeval' can generate receiver operating characteristic (ROC) curves, precision-recall (PR) curves, calibration curves, and PR gain curves. 'MLeval' accepts a data frame of classification probabilities and ground truth labels, or it automatically parses the results of the 'Caret' training function for repeated cross-validation, then selects the best model and analyzes the results. 'MLeval' generates a variety of evaluation metrics with confidence intervals, Comprehensive RArchive Network (CRAN), 2020-02-12 06:40:02UTCMlogit, 1.1-1, "Maximum Likelihood Estimation of Random Utility Discrete Choice Models. Software described in Croissant (2020)<doi:10.18637 / jss.v095.i11> , see Train (2009) for the underlying method<doi:10.1017 / CBO9780511805271> . ”, Comprehensive R ArchiveNetwork (CRAN), 2020-10-02 12:12:05UTC

[0140] Modeltools, 0.2-23, "A collection of tools for working with statistical models. The functionality is experimental and the user interface may change in the future. Documentation is sparse, but packages 'coin' and 'party' have working examples. However, if you are interested in implementing the idea, we are very interested in discussing this proposal. Contributions are very welcome!", Comprehensive R Archive Network(CRAN), 2020-03-05 11:50:06UTC

[0141] Mvtnorm, 1.1-1, "Computes multivariate normal and t distribution probabilities, quantiles, random deviations, and densities.", Comprehensive R Archive Network (CRAN), 2020-06-09 15:50:02UTC

[0142] NCmisc, 1.1.6, "A set of convenient functions. Includes a versatile one-line progress bar, a one-line function timer with detailed output, time delay functions, text histograms, object previews, Comprehensive R Archive Network (CRAN) package search, a simpler package installer, Linux command installation check, flexible mode functions, top functions, simulation of related data, etc.", Comprehensive R Archive Network (CRAN), 2018-11-12 00:00:03UTCParty, 1.3-6, "A computational toolbox for recursive partitioning. The core of the package is ctree(), which implements conditional inference trees, embedding tree-structured regression models into well-defined conditional inference steps. This nonparametric class of regression trees is applicable to all types of regression problems, including nominal, ordinal, numeric, censored, and arbitrary measurement scales for multivariate response variables and covariates. Based on conditional inference trees, cforest() provides an implementation of Breiman's random forest. The function mob() implements a recursive partitioning algorithm based on parametric models (such as linear models, GLMs, or survival regression), using parameter instability tests for split selection. Extensible functions for visualizing tree-structured regression models are available. The method is described in Hothorn et al. (2006)<doi:10.1198 / 106186006X133933> ,Zeileis et al. (2008)<doi:10.1198 / 106186008X319331> and Strobl et al. (2007)<doi:10.1186 / 1471-2105-8-25> . ”, Comprehensive R Archive Network (CRAN), 2021-02-0812:20:02UTC

[0143] Partykit, 1.2-12, "A toolkit with infrastructure for representing, summarizing, and visualizing tree-structured regression and classification models. This unified infrastructure can be used to read / force tree models from different sources (e.g., 'rpart', 'RWeka', 'PMML'), generating objects that share print() / plot() / predict() method functionality. In addition, conditional inference trees (ctree()) and model-based recursive partitioning (mob()) from the 'party' package are provided based on the new infrastructure." Package description by Hothorn and Zeileis (2015)<https: / / jmlr.org / papers / v16 / hothorn15a.html> Published. ”, Comprehensive R Archive Network (CRAN), 2021-02-08 15:10:03UTC

[0144] PerformanceAnalytics, 2.0.4, "A collection of econometric functions for performance and risk analysis. In addition to standard risk and performance measures, the package aims to help practitioners and researchers leverage recent research on the analysis of abnormal return streams. In general, the functions are tested on regular-sized return (rather than price) data, but most functions also work with irregular return data, and an increasing number of functions will work with P&L or price data where possible.", Comprehensive RArchive Network (CRAN), 2020-02-06 12:10:11 UTC

[0145] plotROC, 2.2.1, "Most ROC curve plots obscure the cutoff values ​​and hinder the interpretation and comparison of multi-day curves. This attempts to address these shortcomings by providing plotting and interactive tools. Functions for generating interactive ROC curve plots are provided for web use, as well as printed versions. A Shiny application implementing these functions is included.", ComprehensiveRArchive Network (CRAN), 2018-06-23 07:52:56 UTC

[0146] plyr, 1.8.6, "A set of tools for solving a common problem: the need to break a large problem into manageable parts, operate on each part, and then put all the parts back together. For example, one might want to fit a model to every spatial location or time point in a study, summarize data by panel, or collapse high-dimensional arrays into simpler summary statistics. The development of plyr was generously supported by Becton Dickinson.", Comprehensive R Archive Network (CRAN), 2020-03-03 14:40:02 UTC

[0147] png, 0.1-7, The package provides an easy way to read, write, and display bitmap images stored in the PNG format. It can read and write raw vectors in files and in memory. , Comprehensive R Archive Network (CRAN), 2013-12-03 22:25:05UTC

[0148] pROC, 1.17.0.1, Tools for visualizing, smoothing, and comparing receiver operating characteristics (ROC curves). The (partial) area under the curve (AUC) can be compared with statistical tests based on the U statistic or bootstrap. Confidence intervals for (p)AUC or ROC curves can be calculated, Comprehensive R Archive Network (CRAN), 2021-01-13 14:30:02UTC

[0149] psych, 2.0.12, a general toolkit for personality, psychometric theory, and experimental psychology. Functions are primarily for factor analysis, principal components analysis, cluster analysis, and reliability analysis, although others provide basic descriptive statistics. Item response theory is done using factor analysis with four-item and multinomial correlations. Functions for analyzing data at multiple levels include within- and between-group statistics, including correlations and factor analysis. Functions for modeling and testing specific item and test structures are included. Several functions serve as useful front-ends to structural equation modeling. Basic graphics are used to create path diagrams, factor analyses, and graphical displays of structural equation models. Some functions are written to support psychometric theory books as well as personality research publications. For more information, see<https: / / personality-project.org / r / > Web page, Comprehensive R Archive Network (CRAN), 2020-12-16 16:10:03UTC

[0150] readxl, 1.3.1, "Import excel files into R. via the embedded 'libxls' C library<https: / / github.com / libxls / libxls> Support for '.xls' via embedded 'RapidXML' C++ library<http: / / rapidxml.sourceforge.net> Supports '.xlsx'. Runs on Windows, Mac, and Linux without external dependencies. ", Comprehensive R Archive Network (CRAN), 2019-03-13 16:30:02UTC

[0151] reshape2, 1.4.4, "Flexibly reshape and aggregate data using only two functions, melt and 'dcast' (or 'acast'), to flexibly reshape and aggregate data.", Comprehensive R Archive Network (CRAN), 2020-04-09 13:50:02 UTC

[0152] ROCR, 1.0-11, "ROC plots, sensitivity / specificity curves, lift plots, and precision / recall plots are popular examples of visualizations of trade-offs for specific performance metrics. ROC is a flexible tool for generating two-dimensional performance curves parameterized by cutoffs by freely combining two of 25 performance metrics (new performance metrics can be added using a standard interface). Curves from different cross-validation or bootstrap runs can be averaged by different methods, and the variability over the runs can be visualized using standard deviation, standard error, or boxplots. It supports averaging curves from different cross-validation or bootstrap runs and showing the variability by standard deviation or standard error. Parameterizations can be visualized by printing the cutoffs at the corresponding curve positions or by coloring the curves according to the cutoff. All components of the performance plots can be quickly adjusted using a flexible parameter scheduling mechanism. Despite its flexibility, ROCR is easy to use, with only three commands, and all optional parameters have sensible defaults.", Comprehensive R Archive Network (CRAN), 2020-05-02 14:50:05UTC

[0153] Rpart, 4.1-15, "Recursive partitioning for classification, regression, and survival trees. An implementation of much of the functionality of the 1984 book by Breiman, Friedman, Olshen, and Stone.", Comprehensive R Archive Network (CRAN), 2019-04-12 14:32:39 UTC

[0154] Sandwich, 3.0-0, "Object-oriented software for model-robust covariance matrix estimators. Starting with the basic robust Eicker-Huber-White sandwich covariance method, including: heteroskedasticity-consistent (HC) covariance for cross-sectional data, heteroskedasticity- and autocorrelation-consistent (HAC) covariance for time series data (e.g. Andrews' kernel HAC, Newey-West, and WEAVE estimators), clustered covariance (one-way and multi-way), panel and panel-corrected covariance, gradient outer product covariance, and (clustered) bootstrap covariance. All methods apply to (generalized) linear model objects fitted by lm() and glm(), and can be adapted to other classes via S3 methods. See Zeileis et al. (2020) for details.<doi:10.18637 / jss.v095.i01> , Zeileis et al. (2004)<doi:10.18637 / jss.v011.i10> and Zeileis et al. (2006)<doi:10.18637 / jss.v016.i09> . "Comprehensive R Archive Network (CRAN), 2020-10-02 10:40:02UTC

[0155] Stabs, 0.6-4, "Resampling procedures are used to assess the stability of selected variables, providing additional finite sample error control for high-dimensional variable selection procedures such as Lasso and Boosting. This implements the standard stability selection (Meinshausen & Buhlmann, 2010,<doi:10.1111 / j.1467-9868.2010.00740.x> ) and complementary pair stability selection with improved error bounds (Shah & Samworth, 2013,<doi:10.1111 / j.1467-9868.2011.01034.x> ). The package can be combined with any user-specified variable selection method. ”, ComprehensiveR Archive Network(CRAN), 2021-01-2910:00:02UTC

[0156] Strucchange, 1.5-2, "Testing, monitoring, and identifying structural change in (linear) regression models. Strucchange features tests / methods from the generalized volatility test framework and the F-test (e.g., Chow test) framework. This includes methods for fitting, plotting, and testing volatility procedures (e.g., CUSUM, MOSUM, recursive / moving estimation) and F-statistics, respectively. The volatility procedure can be used to monitor data inputs online. Finally, breakpoints and confidence intervals in regression models with structural change can be estimated. Emphasis is on methods for visualizing data." Comprehensive R Archive Network (CRAN), 2019-10-12 18:35:49 UTC

[0157] Survival, 3.2-7, "Contains core survival analysis routines, including definitions of the Surv object, Kaplan-Meier and Aalen-Johansen (multistate) curves, Cox models, and parametric accelerated failure time models.", ComprehensiveRArchive Network (CRAN), 2020-09-28 08:20:02UTC

[0158] VIM, 6.1.0, "New tools for visualizing missing and / or imputed values ​​are introduced, which can be used to explore the structure of data and missing and / or imputed values. Based on this structure of missing values, corresponding methods can help identify the mechanism generating missing values ​​and allow exploration of data containing missing values. In addition, the quality of imputation is intuitively explored through a variety of univariate, bivariate, multivariate, and multivariate plotting methods. A graphical user interface available in a separate package VIMGIII allows easy handling of the implemented plotting methods.", Comprehensive R Archive Network(CRAN)2021-01-19 17:00:13UTC

[0159] Xlsx, 0.6.5, provides R functions to read / write / format Excel 2007 and Excel 97 / 2000 / XP / 2003 file formats, Comprehensive R Archive Network (CRAN), 2020-11-10 15:00:02

[0160] Xts, 0.12.1, provides unified handling of different temporal data classes in R by extending zoo, maximizing native format information preservation and allowing user-level customization and extension while simplifying cross-class interoperability, Comprehensive R Archive Network (CRAN), 2020-09-09 16:40:03UTC

[0161] Zoo, 1.8-8, "An S3 class for observations with fully ordered indices. Specialized for numeric vectors / matrices and factors of irregular time series. A key goal of Zoo is to be independent of specific index / date / time classifications, consistent with the ts class and base R by providing methods to extend the standard generics.", Comprehensive R Archive Network (CRAN) 2020-05-02 16:14:00 UTC

[0162] Fig. 9 It is schematically shown how the computer-implemented method as an artificial intelligence model functions.

[0163] like Fig. 9 As shown, the input can be extracted features or parameters from sensor data and game results. The output can be a predictor of the likelihood of a person suffering from different sizes of nerve damage. The classification model can save data patterns of healthy subjects and patients with peripheral neuropathy, which are learned from previously acquired data sets. The workflow can be that the model attempts to draw conclusions about the likelihood of a person suffering from peripheral neuropathy from the input values. The calculation process can occur in the data processing device 14. The application model of the learned data pattern can be a support vector machine (SVM), a random forest (RF), an artificial neural network (ANN), etc.

[0164] Fig.10 is an overview of an exemplary embodiment of the system 10. Area A includes a shoe having a sensor-equipped gaming insole having eight embedded pressure sensors 20 ( Fig.10 not shown). Fig.10 The upper middle image of shows the positioning of these sensors at the calcaneus at the heel, the arch, the metatarsal bones (Met) 1 / 3 / 5, the toes and the phalanges at the big toe of the forefoot. The insole acts as a steering unit and is connected to a control unit for real-time data transmission, for example, wirelessly (e.g., Bluetooth) to a gaming application, which runs on a tablet shown in area B. This setup allows a person to play games individually by adjusting the plantar pressure value. Each gaming session includes eight calibration steps and four games, namely Balloon Flight (BF), Apple Catch (AC), Cross Pressure (CP), and Island Jump (IJ), as shown in the middle of the middle image. Subsequently, the data may be uploaded as a data packet to a server shown in area C. A person may receive a briefing of his or her performance, including a spider graph, score, level, and ability (beginner, average, advanced, or expert), as shown in the lower middle image. In addition, the physician may implement further data visualization and data analysis to indicate the performance of the person in different games compared to the maximum completion level.

[0165] Fig.11An example correlation between nerve conduction velocity and features derived from sensor data collected from people playing the corresponding games is shown. TCL stands for "task combination for left foot" and TCR stands for "task combination for right foot". The equation and adjusted R were calculated by a simple linear regression model. 2 Pearson and Spearman correlations were used for the analysis of normally and non-normally distributed data, respectively.

[0166] Fig.12 Peripheral neuropathy classification and game-based findings in subgroups are shown. Region A shows an illustration of hypothesis-driven key capabilities between people without (w / o) and with symmetrical PNP. Region B shows a visualization of the game performance of two study subjects (healthy individuals without PNP, symmetrical PNP patients). Region C shows representative game features associated with the presence of PNP. Region D shows an assessment of the predictive ability of the trained classification model to distinguish between feet with PNP and healthy feet (w / o PNP). Fig.12 The abbreviations used in the study are: PNP, peripheral neuropathy; AC, catching the apple game; BF, balloon flight game; CP, cross-pressure game; IJ, island hopping game; TCL, left-foot task combination; TCR, right-foot task combination; GLMNET, lasso and elastic-net regularized generalized linear model; AUC-ROC, area under the receiver operating characteristic curve.

[0167] Fig.13 Nerve injury patterns based on clinical examination and correlation with game-based findings are shown. Region A shows clinical examination results for small and large fiber neuropathy. Region B shows game features significantly associated with nerve fiber injury subclassifications. Region C is a visualization of the multi-classification AUC-ROC curves of the PNP subclassifications (small / large / mixed nerve fiber dysfunction). Fig.13 The abbreviations used in the study are: PNP, peripheral neuropathy; SFN, small-fiber neuropathy; LFN, large-fiber neuropathy; AC, apple-catching game; BF, balloon-flying game; CP, cross-pressure game; IJ, island-hopping game; TCL, left-foot task combination; TCR, right-foot task combination; AUC-ROC, area under the receiver operating characteristic curve.

[0168] Fig.14 An exemplary embodiment of a calibration procedure for calibrating a foot pressure sensor is shown. Fig.14 Part A shows example actions to be performed by a person for a calibration procedure. Fig.14 Part B details the exemplary calibration steps of Part A. Part C illustrates an applicable normalization algorithm.

[0169] Fig.15An exemplary embodiment of how to extract features for a PNP classification model to perform game feature analysis of a subgroup is shown. NCS was performed in 37 patients (subgroup 1), all of whom were diagnosed with diabetes and had normal cognitive status (MoCA≥26). The distribution of age, weight, BMI, and neuropathy scores in this subgroup was similar to that of the entire group. The gender distribution in the subgroup was slightly biased towards female patients. The patient's game performance was significantly correlated with the sensory and motor conduction velocity (NCV) and amplitude of the lower limbs. Feature extraction of exemplary different parameters from four games was set for each foot. Overall, 277 independent exemplary game features showed significant associations with NCS parameters. Fig.11 Eight representative exemplary features are shown. The strongest correlation can be found between the motor nerve conduction amplitude of the right foot and the maximum execution time in Task 8 of the island jumping game (R = 0.65, p < 0.001, adjusted R2 = 0.36, Fig.11 H section). BF and AC games may contribute further features related to NCV and amplitude of the left foot. However, more features extracted from BF and IJ games may be associated with NCV and amplitude of the right foot. 164 and 113 independent exemplary game features appear to be significantly associated with motor and sensory neural states, respectively. 146 independent exemplary game features appear to be associated with NCV and 131 appear to be associated with amplitude of the nerves of the foot.

[0170] Fig.16 An exemplary embodiment of consideration of key abilities and scoring for a virtual game is represented. For example, of all people (e.g., n=329), a subgroup 2 of age greater than 55 years and normal cognitive status (MoCA≥26) can be generated. There were no significant differences in age, gender, and BMI between patients without (w / o) and with clinical PNP. Individuals with neuropathy had a significantly longer duration of diabetes and more severe neuropathy symptoms. Clinical evidence of PNP was present in 118 left feet and 120 right feet. For all of these feet, at least one site of insensitivity was found during the 10g monofilament test. PNP occurred in both feet of 112 patients. 47 patients had no PNP, and 26% of patients had unilateral PNP.

[0171] According to the present invention, key skills are related to peripheral nerve function, which contribute to successful game performance. As an example, six key abilities are defined: reaction time (understanding; immediate reaction to the task), sensation (fine-tuning of pressure application in subtasks), proficiency (overall completion of each game), muscle strength (lower limbs) (completion with high pressure application), plantar pressure deviation (pressure distribution between left and right feet) and endurance (stability of pressure application in tasks), such as Fig.16Selected or all key abilities showed significant differences between patients without PNP (n=47) and with symmetric PNP (n=112). Example game performance differences between two individuals (without PNP; with PNP; Fig.12 Of the 2622 extracted game features, for example, 50 and 56 independent game features can be used to reveal differences between the “without PNP” and “with PNP” groups. For each game, an illustrative example of comparing and visualizing these gameplay features is shown in Figure 2. Fig.12 of the FDA).

[0172] The prediction model for PNP was trained using five-fold ten-fold repeated cross validation with extracted game features from 173 patients in subgroup 2. For example, 70% of the game dataset was used for model training and 30% for testing. In the retained test set, the obtained lasso and elastic net regularized generalized linear model identified PNP of the left foot with an adjusted accuracy of 76.1% (sensitivity 77.1%, specificity 75.0%, AUC-ROC 0.72, Fig.12 For determining PNP of the right foot, another GLMNET model achieved an accuracy of 81.7% (sensitivity 83.3%, specificity 80.0%, AUC-ROC 0.75, Fig.12 of the Regulations).

[0173] In addition, for example, a multi-classification model was built to distinguish between feet without PNP (w / o) and those with SFN, LFN, and mixed nerve fiber damage. Exemplary game features were extracted from the game dataset for the left and right feet, respectively. The trained gradient boosting model (GBM) achieved a multi-classification AUC of 0.76 and 0.72 for identifying the left and right foot nerve fiber damage patterns, respectively (see Fig.13 Part C of Figure 3). Both models distinguish SFN and LFN well.

[0174] Reference numerals

[0175] 1 Methods

[0176] 2 Getting Started

[0177] 3 Touch your foot to the sensor

[0178] 4 Play the game and collect data

[0179] 5Analyze sensor data and indicate or determine peripheral neuropathy

[0180] 6 End

[0181] 10 System

[0182] 11.12 Foot pressure measurement device

[0183] 13 Gaming equipment

[0184] 14 Data processing equipment

[0185] 20 Pressure sensitive element

[0186] 21 Signal to Data Converter

[0187] 22 Data Interface

Claims

1. A computer-implemented method (1) for indicating whether a person suffers from peripheral neuropathy, It is characterized in that The method (1) comprises loading sensor data from foot pressure sensors (11, 12), which are pressed between the person's feet and the ground when the person plays a predetermined computer game, and controlling the computer game by applying pressure to the pressure sensors through the person's feet, and The sensor data is analyzed (5) by an algorithm, the algorithm being suitable for detecting whether a person suffers from peripheral neuropathy based on the sensor data.

2. The method (1) according to claim 1, characterized in that: The sensor data and game data about the person's game performance are combined and analyzed by algorithms (5).

3. The method (1) according to claim 1 or 2, characterized in that: The sensor data includes data indicative of an amount of pressure applied by a person and / or a timing of pressure applied by a person.

4. The method (1) according to claim 2 or 3, characterized in that: Game data includes the timing and / or accuracy of positioning of an object controlling a game relative to another object of the game.

5. The method (1) according to any one of claims 1 to 4, characterized in that The algorithm is adapted to indicate a person's reaction time, anticipation time, feel, proficiency, endurance, plantar pressure deviation and / or lower extremity muscle strength from sensor data or a combination of sensor data and game data.

6. The method (1) according to any one of claims 1 to 5, characterized in that An algorithm is an artificial intelligence model that is trained to indicate whether a person has peripheral neuropathy based at least on pressure data.

7. A data processing device (14) having a storage device and a processor, wherein the storage device comprises instructions which, when executed by the processor, cause the processor to perform the method of any one of claims 1 to 6.

8. A computer program comprising instructions which, when executed by a computer, cause the computer to execute the method (1) as claimed in any one of claims 1 to 6.

9. A system (10) for indicating whether a person suffers from peripheral neuropathy, It is characterized in that The system comprises a gaming device (13) for playing a computer game, at least two foot pressure sensors (11, 12) connectable to the gaming device (13) in a control signal transmission manner, and The data processing device (14) as claimed in claim 7, wherein: The data processing device (14) is adapted to load sensor data from the plantar pressure sensors (11, 12), the gaming device (13) or another storage device of the system.

Images