You are on page 1of 4

Introduction to Artificial Neural Networks (ANNs) and a method of how a classification problem can be solved using ANNs The

problem we are facing is the classification of Gait patterns from the DGI tests. The solution for the problem of classification can be done by using Artificial Neural Networks. The functions of ANNs are similar to how brain functions biologically. A brain consists of nerve cells or neurons which receives stimulus and triggers a response from the neuron. This neuron triggers another neuron and eventually forms a neural activation chain which results in some response of the body like identification of sound, recollecting a scene from memory etc. The way ANNs represented in computer is similar. ANNS have many neurons but are divided into layers. Input layer which receives the initial input or stimulus .The input layers is connected to a hidden layer and hidden layer is connected to another hidden layer or to an output layer. There can be any number of hidden layers. But typically the number of hidden layers is kept limited in any ANN. Training of a neural network is necessary in order to get a NN learning the task successfully. The training data has to be collected and should include the inputs to each neuron and also the anticipated output to the output neurons. More number of examples can be given so that learning procedure would be efficient. When the network is created it has some pre determined weights. The most common training or learning technique used is back propagation. This technique takes the input values and pulses with the network and produces the output values. This output values are compared with the expected outputs, if the output are unlikely then the weights are modified based upon the error involved and the learning rate . the value of ranges from 0.3 to 0.7 and indicates the percentage of error used to modify the weights. Weights in the NN are typically adjusted backwards. ANNs are mostly used in Classification problems. The problem is one where certain number of classes is identified. On the basis of input data NN tells to which class the input data most likely belong to. If dealing with the classification problem the number of output nodes is determined by the number of classes and each output node represents one class. For example in our problem of classifying the gait patters from 8 DGI tests. The output layer of the neural network has to have 8 nodes, each node representing each test. In this way the value of the output node tells the likelihood of the input sample falling into each of these tests, determined by the neural network. And in the similar way if we want to detect a fall, the output layer has to have nodes which represent the specific threshold values. The threshold values can be calculated by recording the ACC X, ACC Y, ACC Z (Ax, Ay Az), Gyro X, Gyro Y and Gyro Z (x, y, z) by placing the WGAS at the patients T4 position or at the nape area to detect any fall.

For example if we want to detect a fall to have 100% sensitivity from three thresholds out of four, we can have ANN with the 4 output nodes. And similarly if we want to detect to have 100% sensitivity from four thresholds out of seven, ANN should have seven output nodes. Depending upon the output values of the NN we can observe how many outputs have exceeded the expected output (in our case these are threshold values generated, Th1: accelerometer threshold, Th2: angular velocity threshold, Th3: angular acceleration threshold, Th4: Total angular change threshold, Th5: angular change in the X-axis threshold, Th6: angular change in the Y-axis threshold, Th7: angular change in the Z-axis threshold).In this way ANNs are useful to classify a specific class from the set of different classes. But however if the database of the patients becomes very large I.e. more number of example are provided as inputs to the ANN employing back propagation , the training or the learning procedure becomes very late and have serious effect on the learning rate . The larger value of indicates that weights of the network need greater adjustment. The learning rate of the network is important because NN is supposed to learn all of the examples in the training data. By over adjusting the weights can cause NN to unlearn some of the examples. That is why has to be decreased for effective learning without counter acting the previous progress. This is one of the disadvantages we can see with the ANN with back propagation if the training data is very large. We always have to keep track of the learning rate for better output results. So there is a need to come up with a good Neural Network topology to deal with the problem mentioned above.

Figure 1 Comparing DGI Test 1 and 2 Table 1 outliers table for DGI Test 1 and 2

Test # Test 1 Test 1 Test 2 Test 2

Sensor Placement Nape T4 Nape T4

Total outliers (All 6 axes) 67 208 189 510

Only ACC X outliers 0 0 13 (Higher side) 26 (Lower side)

Only ACC Y outliers 0 7 (Higher side) 12 (Lower side) 29 (Lower side) and 39 (Higher side)

For DGI test1 at nape the difference between the median values of ACC Y and X is in between 0.057 to 0.089 and at t4 is in between 0.009 to 0.021.so Difference of median values at nape is nearly 5 times of T4 value. That is why we observe box plots of Acc X and Y are far from each other at Nape but at T4 they seem to be very close like almost equal. Considering the same trend for DGI test 2 median differences of ACC Y and X is in between 0.089 to 0.109 and at t4 is in between -0.012 to -0.003. We observe that median difference at nape is large comparative to test 1 but interesting thing is that all medians of ACC Y falls behind ACC X values at T4 in DGI test 2 and that is why we see negative values (-0.012 to -0.003)

Figure 2 Comparison of DGI Test 3 and 4

Acc Y values are tighter at DGI Test 4 at T4 compared to Nape. The STDEV of ACC Y at T4 in test 3 is 0.02744 and at Nape is 0.045777.In test 4 the stdev of ACC Y at t4 is 0.029204 and at Nape is 0.070289.We observe some increase in the STDEV of ACC Y at Nape in test 4 but not at T4 in test 4 (tighter box plot has smaller STDEV).So I guess the STDEV of ACC Y at nape in test 4 can be set as a rule to differentiate DGI test #3 and Test #4. From the data and STDEV values I agree with your point.

Figure 3 Comparion of DGI Test 6 and 7

You might also like