CN107157450A - Quantitative estimation method and system are carried out for the hand exercise ability to patient Parkinson - Google Patents

Abstract

The present invention provides a method and system for quantitatively evaluating the hand movement ability of Parkinson's patients, using a wearable myoelectric sensor to capture the hand surface electromyography signal of the detected person when performing a specified action, based on the trained The motion ability classifier corresponding to the specified action extracts time-domain features and frequency-domain features from the surface electromyographic signal, as well as features related to the completion of the specified action, to quantitatively evaluate the hand motion ability of the detected person. The method and system can more objectively and accurately evaluate the hand movement ability of Parkinson's patients.

Description

Method and system for quantitative evaluation of hand motor ability of Parkinson's patients

technical field

The invention relates to the quantitative evaluation of motor ability, in particular to a quantitative evaluation method and system for Parkinson's patient's hand motor ability.

Background technique

The motor symptoms of patients with Parkinson's Disease (PD) mainly include slowness of movement, resting tremor, and muscle stiffness. At present, the detection and evaluation of the motor ability of PD patients mainly relies on the PD Rating Scale (Unified Parkinson's Disease Rating Scale, UPDRS) provided by the International Sports Association for grade evaluation. For example, the hand motor ability of Parkinson patients is usually divided into It is graded from 0 to 4. Grade 0 is equivalent to the hand motor ability of normal people. The higher the grade, the worse the hand motor ability. However, this evaluation method is usually affected by the scoring physician's operating experience and the state and emotion of the person being tested during the evaluation, so the evaluation results are still not objective and accurate enough.

Contents of the invention

Therefore, the object of the present invention is to overcome the defects of the above-mentioned prior art, and provide a method and system for quantitatively evaluating the hand motor ability of a Parkinson's patient by using the surface electromyographic signal of the hand.

The purpose of the present invention is achieved through the following technical solutions:

On the one hand, the present invention provides a method for quantitatively evaluating the hand motor ability of Parkinson's patients, comprising:

Step 1, collect the surface electromyography signal when the subject performs a specified action via the electromyographic sensor placed on the subject's hand;

Step 2, using the pre-trained Parkinson's patient motor classifier corresponding to the specified action to judge the grade of the hand motor ability of the detected person according to the collected surface electromyographic signals;

The features used to train the Parkinson's patient's motor ability classifier corresponding to the specified action include extracting time-domain features and frequency-domain features from the surface electromyography signal and obtaining information related to the completion of the specified action based on the surface electromyography signal. Characteristics.

In the above method, the feature related to the completion of the specified action may include at least one of the following: the maximum electromyographic signal when the specified action is completed, and the time taken to complete the specified action.

In the above method, the specified action may include at least one of the following: making a fist and pointing fingers.

In the above method, it may also include the step of training a Parkinson's patient motor ability classifier corresponding to the specified action, which includes:

a) receiving a plurality of surface myoelectric signals collected by a myoelectric sensor worn by a plurality of different degrees of Parkinson's patients and normal people when performing the specified action as a training data set;

b) extracting time-domain features and frequency-domain features from each surface electromyography signal and obtaining features related to the completion of the specified action based on the surface electromyography signal;

C) select the feature for training the athletic ability classifier based on the information gain of each feature for the training data set, wherein the information gain of each feature for the training data set is the experience entropy of the training data set and given this feature The difference between the empirical conditional entropy of the training dataset under condition ;

d) Training the athletic ability classifier based on the selected features.

In the above method, when the specified action is pointing fingers, a sequential minimal optimization model can be used as the athletic ability classifier; when the specified action is fist clenching, a J48 classifier can be used as the athletic ability classifier.

In the above method, the empirical entropy of the training data set represents the uncertainty of classifying the training data set, which can be calculated by the following formula:

Among them, D represents the training data set, H(D) represents the experience entropy of the training data set D, n represents the training data set D is divided into n categories, and p _i represents the probability of dividing the data into the i-th category.

In the above method, the empirical conditional entropy of the training data set under the condition of a given feature represents the uncertainty of classifying the training data set under the condition of a given feature, which can be calculated by the following formula:

Where A represents a certain feature, H(D|A) is the experience entropy of the training data set D under the condition of given feature A, p _i (D|A) means that the data is divided into The probability of class i.

In yet another aspect, the present invention provides a system for quantitatively evaluating the hand motor ability of Parkinson's patients, comprising:

The acquisition device is used to collect the surface electromyography signal of the subject when the subject performs a specified action via the myoelectric sensor placed on the subject's hand;

The detection device is used to use the pre-trained Parkinson's patient movement ability classifier corresponding to the specified action to judge the grade of the hand movement ability of the detected person according to the collected surface electromyographic signal;

The features used to train the Parkinson's patient's motor ability classifier corresponding to the specified action include extracting time-domain features and frequency-domain features from the surface electromyography signal and obtaining information related to the completion of the specified action based on the surface electromyography signal. Characteristics.

In the above system, the features related to the completion of the specified action may include at least one of the following: the maximum electromyographic signal when the specified action is completed, and the time taken to complete the specified action.

In the above system, the specified action may include at least one of the following: making a fist and pointing fingers.

The system described above may also include a training device for training a Parkinson's patient motor capacity classifier corresponding to a specified action, which is configured to:

a) receiving a plurality of surface myoelectric signals collected by a myoelectric sensor worn by a plurality of different degrees of Parkinson's patients and normal people when performing the specified action as a training data set;

b) extracting time-domain features and frequency-domain features from each surface electromyography signal and obtaining features related to the completion of the specified action based on the surface electromyography signal;

C) select the feature for training the athletic ability classifier based on the information gain of each feature for the training data set, wherein the information gain of each feature for the training data set is the experience entropy of the training data set and given this feature The difference between the empirical conditional entropy of the training dataset under condition ;

d) Training the athletic ability classifier based on the selected features.

In the above system, when the specified action is pointing fingers, a sequential minimal optimization model can be used as the athletic ability classifier; when the specified action is fist clenching, a J48 classifier can be used as the athletic ability classifier.

Compared with the prior art, the present invention has the advantages of:

Use the wearable electromyographic sensor to capture the surface electromyographic signal of the subject's hand to extract time-domain features and frequency-domain features from the surface electromyographic signal, as well as features related to the completion of the specified action to analyze the subject's hand. Quantitative assessment of the hand movement ability of Parkinson's patients can be performed more objectively and accurately.

Description of drawings

Embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

FIG. 1 is a schematic flowchart of a method for quantitatively evaluating the hand motor ability of a Parkinson's patient according to an embodiment of the present invention;

Figure 2(a) is a schematic diagram of the comparison of the recognition accuracy of each classifier corresponding to the "pointing finger" action;

Figure 2(b) is a schematic diagram of the comparison of the recognition accuracy of each classifier corresponding to the "clench fist" action;

Fig. 3 (a) is the SMO classifier recognition performance schematic diagram that utilizes the signal recognition feature training of surface electromyographic signal for " pointing to " action;

Figure 3 (b) is a schematic diagram of the J48 classifier recognition performance training using the signal recognition feature of the surface electromyography signal for the "clench fist" action;

Fig. 4 (a) is a schematic diagram of classifier recognition performance corresponding to the "finger" action trained by the method according to the present invention;

Fig. 4 (b) is a schematic diagram of classifier recognition performance corresponding to the action of "clenching a fist" trained according to the method of the present invention;

Figure 5(a) is a schematic diagram of the comparison of the recognition results of the classifiers corresponding to the "finger-to-finger" action trained with different features;

Figure 5(b) is a schematic diagram of the comparison of the recognition results of the classifier corresponding to the "fist" action trained with different features.

detailed description

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below through specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

The "pointing" and "fist" gestures are two commonly used international standard indicators for assessing the hand motor ability of Parkinson's patients. In the following, the two specified actions will be taken as examples to describe how to evaluate the hand motor ability of Parkinson's patients in combination with specific embodiments. However, it should be understood that the embodiments of the present invention are not limited to specific sports actions, and the principles of the present invention may also be applied to other actions or the assessment of the athletic ability of other parts.

Figure 1 shows a method for quantitatively evaluating the hand motor ability of a Parkinson's patient according to an embodiment of the present invention, which includes collecting the data of the subject via a myoelectric sensor placed on a relevant part of the subject's hand. The surface electromyographic signal (step 1) when performing the specified action; the Parkinson's patient's motor ability classifier corresponding to the specified action is used to judge the hand movement of the detected person according to the collected surface electromyographic signal The level to which the ability belongs (step 2). The motor capacity classifier of Parkinson's patients corresponding to the specified action is trained on the basis of sample data extracted from a large number of Parkinson's patients of different degrees and some normal people when performing the specified action.

More specifically, in step 1, the myoelectric sensor is placed on a designated position of the subject, such as an arm and other relevant parts, and then the subject is asked to perform a designated action, such as making a fist or pointing fingers. Here, any type of myoelectric sensor capable of collecting human surface myoelectric signals can be used. Preferably, a wearable electromyographic instrument can be used, for example, the armband MYO produced by the Canadian startup company ThalmicLabs is used below to detect the surface electromyographic signal during hand movement. It is a gesture control armband that can communicate with other Electronics make wireless connections and data transfers. When the subject is making a fist and pointing fingers, the muscles above the forearm are mainly active to generate corresponding myoelectric signals, so MYO can be placed on the forearm of the subject. The electromyographic sensor sends the collected surface electromyographic signals to corresponding signal analysis and processing devices through wired or wireless transmission, for example, mobile terminals such as tablets or mobile phones, desktop computers, and devices capable of processing surface electromyographic signals. any other computing device.

Preferably, the computing device responsible for processing the surface electromyography signals can also perform a certain degree of preprocessing on the received surface electromyography signals, such as filtering out noise or performing dimensionality reduction processing, such as performing processing on signals from multiple electromyography sensors. Synthesis to simplify the computational complexity as much as possible and reduce the consumption of computing resources. Taking _MYO as an example, it consists _of eight sensor chips, which can collect eight _- dimensional _myoelectric signals of the patient's _forearm in _one week _. , a ₈ to represent. When the subject makes a "fist" and "fingers", the muscles on the surface of the forearm will be activated. For example, when the patient makes a fist, the muscles around the forearm are in a tense state at the same time, that is, the activation state, which causes the myoelectric signals of the eight myoelectric sensors on the electromyography instrument to increase simultaneously. When the patient's arm is opened, the muscles of the upper forearm are in a relaxed state at the same time, or called an inactive state, resulting in the reduction or disappearance of the myoelectric signals of the eight myoelectric sensors. The collected eight-dimensional myoelectric signal has a correlation, so the method of synthesizing acceleration can be used to synthesize the collected myoelectric signal, such as formula (1) Thus, the synthesized EMG signal sample is obtained. While using synthetic EMG signals to reduce the computational complexity, it also allows the wearer to ignore the direction in which the EMG is worn, and can wear the EMG composed of eight EMG patches at will. It should be understood that the above preprocessing is a preferred method without any limitation, and the quantitative evaluation of the hand movement ability can also be realized by using the surface electromyographic signal collected by a single electromyographic sensor. The surface electromyographic signal mentioned below no longer specifically refers to the electromyographic signal collected by a single electromyographic sensor or the synthesized electromyographic signal, because the processing methods of the two are similar.

Continuing to refer to Fig. 1, in step 2, after obtaining the surface electromyographic signal when performing the specified action, extract the features required by the Parkinson's patient's motor ability classifier corresponding to the specified action, and then extract the features It is provided to a pre-trained classifier to judge the class to which the detected person's hand motor ability belongs. For each specified action, there is a pre-trained Parkinson's patient's motor ability classifier corresponding to it. For the method described below, the classifier corresponding to the fisting action will be referred to as the fisting classifier for short, and the classifier corresponding to the fingering action will be The classifier of is referred to as the index classifier. The training methods of these two classifiers are similar, so the following only uses the fist classifier as an example to illustrate the training method of the classifier.

During training, a large number of Parkinson's patients and some normal people of different degrees are firstly allowed to wear myoelectric sensors and perform fisting movements, so as to collect a large number of surface electromyography signals as a basic training data set. When training a classifier, the key is to determine which features of the data are used for training, and these features will directly affect the accuracy of the classification results. Usually, time-domain and frequency-domain analysis can be performed on surface EMG signals, from which multiple time-domain features and frequency-domain features can be extracted to train a classifier. These time-domain and frequency-domain features reflect the characteristics of the EMG waveform itself. In a preferred embodiment of the invention, in addition to using time-domain and frequency-domain features, features related to the completion of specified actions are also used. This is because the time and strength used by PD patients with different degrees to complete a specified action are quite different. Therefore, the characteristics related to the completion of the specified action can also better distinguish the hand motor ability. sex. In order to improve the accuracy of the classifier itself, in the embodiment of the present invention, some optimal features are selected from the above multiple features based on information gain to train the classifier. The process of training a classifier mainly includes:

(1) Extract time domain features and frequency domain features from surface EMG signals

Surface electromyogram signal (SEMS) is a bioelectrical signal emitted when neuromuscular activity is recorded from the surface of the human skin through sensors or electrodes, and is noninvasive. The amplitude of the surface electromyography signal is random and can usually be expressed as a quasi-Gaussian distribution function. Usually digital analysis is the main means of processing surface electromyographic signals, including time domain analysis and frequency domain analysis. Time-domain analysis regards the surface EMG signal as a function of time, and obtains some statistical characteristics of the EMG signal through analysis. For example, common time-domain features include the average value of the EMG signal in the time domain, the maximum value, variance, standard deviation, mode, zero-crossing times, mean square value, mean-crossing rate, etc. Frequency domain analysis is to convert time domain signals into frequency domain signals through Fourier transform. Common frequency domain features include peak value, frequency component, DC component, energy, shape mean, shape standard deviation, shape skewness, shape kurtosis, amplitude Mean, amplitude standard deviation, amplitude skewness, amplitude kurtosis, median frequency, mean frequency, frequency range, highest peak frequency, highest peak amplitude, and more. The above time-domain and frequency-domain features may also be collectively referred to as signal identification features.

(2) Obtain features related to the completion of the specified action based on the surface electromyographic signal

Features related to the completion of the specified action (may be referred to as action features for short) include the maximum electromyographic signal when the specified action is completed, the time taken to complete the specified action, and the like. With the aggravation of PD patients, motor symptoms of Parkinson's disease will pause and slow down, which will lead to a decrease in the strength of EMG. In addition, when the motor symptoms of PD are relatively severe, the time taken to complete the action will be longer. Therefore, features related to the completion of specified actions can also effectively distinguish the hand motor ability of different degrees of PD patients.

(3) Select effective features from the above-mentioned multiple signal recognition features and action features based on information gain to train a classifier.

First, the information gain of each feature is calculated. Information gain represents the degree to which the uncertainty of class Y information is reduced by knowing the information of feature X. The information gain of feature A to training data set D is denoted as g(D,A), which is the empirical entropy or information entropy H(D) of training data set D and the empirical condition of training data set D under the given condition of feature A The difference between the entropy H(D|A), namely:

g(D,A)=H(D)-H(D|A)

Among them, the empirical entropy H(D) represents the uncertainty of classifying the data set D, and the empirical conditional entropy H(D|A) represents the uncertainty of classifying the data set D under the given condition of feature A, two The difference between the two is the information gain, which indicates the degree of reduction of the uncertainty of the classification of the data set D due to the feature A. Obviously, for the training data set D, the information gain depends on the feature, and different features often have Different information gain, features with large information gain have stronger classification ability. The empirical entropy H(D) can be calculated, for example, by the following formula:

Among them, n indicates that the training data set D is divided into n categories, p _i indicates the probability of dividing the data into the i-th category, and the greater the entropy H is known from the above formula, the greater the uncertainty of the classification. The empirical conditional entropy H(D|A) can be calculated by the following formula, for example:

Among them, n indicates that the training data set D is divided into n categories, p _i (D|A) indicates the probability of classifying the data into the i-th category when the feature A is known, and H(D|A) represents that the feature A is given The degree of uncertainty in classifying the data set D under certain conditions.

After the information gain of each feature of PD motor symptoms can be calculated according to the above formula, the information gain of each feature is compared, and those features with larger information gain are selected as the ones used to train the classifier, from the surface electromyographic signals in the training data set These features are extracted as sample data to train the classifier.

In order to further illustrate the effects of the methods according to the above embodiments, the inventors also conducted the following experiments.

First, a classifier is trained using only time- and frequency-domain features extracted from surface EMG signals. For example, use support vector machine SVM (Support Vector Machine), sequential minimal optimization SMO (Sequential minimal optimization) model, decision tree model such as J48 and random forest RF (RandomForest) model to train data of "clenching fist" and "fingering" actions After processing, the experimental results are shown in Figure 2(a) and 2(b). For the "clench fist" action, the recognition accuracy using the J48 classifier is the highest, which can reach 86.2%; for the "finger-to-finger" action, the recognition accuracy using the SMO classifier is the highest, which can reach 63.7%. Among them, the results of precision (precision), recall rate (Recall) and F-score of the level (ie, 0-4 level) identified by the SMO classifier for the "finger-to-finger" action are shown in Figure 3(a). For each level, the three indicators are arranged in order from left to right in the figure. For the "clench fist" action, the results of using the J48 classifier to identify its class (ie, 0-4 class) of precision, recall and F-score are shown in Figure 3(b). By observing the strength of the EMG signal, it is found that the EMG signal of the finger is weaker than that of making a fist. This is because the muscle activation of the finger is weaker than the muscle activation of the fist, and it is difficult to carry out when the EMG signal is relatively weak. Classification, so the distinction of the electromyographic signals of the fingers is less obvious than that of making a fist, and the accuracy of the recognition of the fingers is not as high as that of making a fist. When the classification level is 0 and 1, the error rate is higher, that is to say, it is not easy to distinguish the finger-to-finger grade of a normal person from that of a patient whose movements are slightly slowed down. A slight slowing of movement occurs. When the classification level is 2 and 3, the accuracy is lower, as shown in Fig. 3(a). This is because when patients do finger alignment tests, patients with grades 2 and 3 have weaker muscle activation, resulting in a small strength of the EMG signal, and subtle changes in the EMG signal with low strength are likely to cause classification confusion the result of. These factors are the main reasons for the high error rate of finger classification.

Next, for the "clench fist" action, use the J48 classifier, and for the "pointing finger" action, use the SMO classifier, but only use the action characteristics of the maximum EMG signal when completing the specified action and the time it takes to complete the specified action to train the classification device. When using the J48 classifier, the classification accuracy of the fist classifier reaches 70.6%, and when using the SMO classifier, the recognition accuracy of the finger classifier reaches 72.5%.

Finally, the first N effective features are selected from the above-mentioned signal recognition features and action features by the above-mentioned method of calculating information gain to train a classifier, and the same classifier model as above is used. For example, when using the J48 classifier, the accuracy of the fist classifier reaches 90.8%. When using the SMO classifier, the finger classifier achieves an accuracy of 82.3%. The recognition accuracy, recall rate and F-score of such a designated action classifier for grades 0-4 of hand ability are shown in Fig. 4(a)-(b), where for each grade, these three indicators are shown in Arranged from left to right. Figure 5(a)-(b) shows the comparison of experimental results using signal recognition features, action features and fusion features, respectively. As shown in Figure 5, compared with only using signal recognition features or only applying to action features, the classification accuracy after using information gain to select effective features (called fusion features in the figure) has been significantly improved.

Although the present invention has been described in terms of preferred embodiments, the present invention is not limited to the embodiments described herein, and various changes and changes are included without departing from the scope of the present invention.

Claims (10)

1. A method for quantifying the hand motor ability of a Parkinson's patient, said method comprising: Step 1, collect the surface electromyography signal when the subject performs a specified action via the electromyographic sensor placed on the subject's hand; Step 2, using the pre-trained Parkinson's patient motor classifier corresponding to the specified action to judge the grade of the hand motor ability of the detected person according to the collected surface electromyographic signals; The features used to train the Parkinson's patient's motor ability classifier corresponding to the specified action include extracting time-domain features and frequency-domain features from the surface electromyography signal and obtaining information related to the completion of the specified action based on the surface electromyography signal. Characteristics. 2. The method according to claim 1, wherein the features related to the completion of the specified action include at least one of the following: the maximum electromyographic signal when the specified action is completed, and the time taken to complete the specified action. 3. The method according to claim 1, wherein the specified action comprises at least one of the following: making a fist, pointing fingers. 4. The method according to any one of claims 1-3, further comprising the step of training a Parkinson's patient's motor capacity classifier corresponding to specified actions, comprising: a) receiving a plurality of surface myoelectric signals collected by a myoelectric sensor worn by a plurality of different degrees of Parkinson's patients and normal people when performing the specified action as a training data set; b) extracting time-domain features and frequency-domain features from each surface electromyography signal and obtaining features related to the completion of the specified action based on the surface electromyography signal; C) select the feature for training the athletic ability classifier based on the information gain of each feature for the training data set, wherein the information gain of each feature for the training data set is the experience entropy of the training data set and given this feature The difference between the empirical conditional entropy of the training dataset under condition ; d) Training the athletic ability classifier based on the selected features. 5. The method according to claim 4, wherein when the specified action is pointing fingers, adopting the sequence minimum optimization model as the athletic ability classifier; when the specified action is making a fist, adopting the J48 classifier as the Ability classifier described above. 6. The method according to claim 4, wherein the empirical entropy of the training data set represents the uncertainty of classifying the training data set, and is calculated with the following formula: <mrow> <mi>H</mi> <mrow> <mo>(</mo> <mi>D</mi> <mo>)</mo> </mrow> <mo>=</mo> <mo>-</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msub> <mi>p</mi> <mi>i</mi> </msub> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mi> </mi> <msub> <mi>p</mi> <mi>i</mi> </msub> </mrow> Among them, D represents the training data set, H(D) represents the experience entropy of the training data set D, n represents the training data set D is divided into n categories, and p _i represents the probability of dividing the data into the i-th category. 7. The method according to claim 6, wherein the empirical conditional entropy of the training data set given a certain feature represents the uncertainty of classifying the training data set given the feature, as follows formula to calculate: <mrow> <mi>H</mi> <mrow> <mo>(</mo> <mi>D</mi> <mo>|</mo> <mi>A</mi> <mo>)</mo> </mrow> <mo>=</mo> <mo>-</mo> <msubsup> <mi>&amp;Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </msubsup> <msub> <mi>p</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>D</mi> <mo>|</mo> <mi>A</mi> <mo>)</mo> </mrow> <msub> <mi>logp</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>D</mi> <mo>|</mo> <mi>A</mi> <mo>)</mo> </mrow> </mrow> Where A represents a certain feature, H(D|A) is the experience entropy of the training data set D under the condition of given feature A, p _i (D|A) means that the data is divided into The probability of class i. 8. A system for quantitative evaluation of the hand motor ability of Parkinson's patients, said system comprising: The acquisition device is used to collect the surface electromyography signal of the subject when the subject performs a specified action via the myoelectric sensor placed on the subject's hand; The detection device is used to use the pre-trained Parkinson's patient movement ability classifier corresponding to the specified action to judge the grade of the hand movement ability of the detected person according to the collected surface electromyographic signal; The features used to train the Parkinson's patient's motor ability classifier corresponding to the specified action include extracting time-domain features and frequency-domain features from the surface electromyography signal and obtaining information related to the completion of the specified action based on the surface electromyography signal. Characteristics. 9. The system according to claim 1, wherein the features related to the completion of the specified action include at least one of the following: the maximum electromyographic signal when the specified action is completed, and the time taken to complete the specified action. 10. The method according to claim 1 or 2, further comprising a training device for training a Parkinson's patient motor capacity classifier corresponding to a specified action, configured to: a) multiple surface electromyographic signals collected by the electromyographic sensors worn by the Parkinson's patients and normal people of different degrees when performing the specified actions as training data sets; b) extracting time-domain features and frequency-domain features from each surface electromyography signal and obtaining features related to the completion of the specified action based on the surface electromyography signal; C) select the feature for training the athletic ability classifier based on the information gain of each feature for the training data set, wherein the information gain of each feature for the training data set is the experience entropy of the training data set and given this feature The difference between the empirical conditional entropy of the training dataset under condition ; d) Training the athletic ability classifier based on the selected features.