Classification of Urine Odour Using Artificial Neural Networks

Abstract

Introduction An odour create a distinctive smell that is caused by specific volatile molecules; it can be either a single chemical compound or a mixture of them. Volatile organic compounds are currently of great interest for non-invasive disease diagnoses such as diabetes [2] and prostate cancers [3]. In this work, the detection of human urine is studied in order to aid incontinent individuals and prevent social embarrassment. Urine odour is known to comprise of more than 200 different types of chemical compounds, including ketones, aldehydes, and esters [1]. Gas Sensing Device The proposed sensing device is compact with three custom-made CMOS non-specific metal oxide semiconductors coated with n-type tungsten trioxide, Pd/Pt doped tin oxide, and p-type copper oxide, respectively. The sensors are interfaced to an integrated Teensy 3.6 microcontroller; and the response data are collected at a 10 Hz sampling rate. This device has been tested via a custom gas testing station with synthetic urine as well as its three individual interference compounds, ammonia, ethyl acetate and acetone. The maximum concentration levels tested in air were 25 ppm, 20 ppm and 200 ppm, respectively, and 5% for synthetic urine. Five repetitions were performed with five concentration steps and three different humidity levels for each gas of interest. The three sensors responded to the target compounds differently with a specific pattern that can be classified through pre-processing and supervised artificial neural networks. Artificial Neural Networks The collected time-series data were first pre-processed to extract the change of sensor response, and auto-ranged to spread the sample variance evenly across all sensors. This step allows the input sensor data to contribute equally to the neural network models and increase the learning speed and prediction accuracy. Two methods were trialled, a shallow neural network with multi-layer perceptrons, and a deep convolutional neural network (CNN). The shallow neural network topology consists of 4 input neurons (no concentration information), 2 hidden layers of 8 neurons each and 4 output neurons for each odour class. This supervised feed-forward backpropagation algorithm is based on the Levernberg-Marquardt approach with sigmoid function for hidden layers. The convolutional neural network has two convolution layers and two pooling layers before the fully connected output layer, which works in the same manner as a shallow neural network. Unlike a typical 2D CNN used in image processing, this 1D network has the kernel size of 1 × n and a single row filtering window tiling down the dataset. In both models, four experiment repeats were used for training and one for testing, until all had been independently tested. Results and Conclusions Confusion matrices are used to demonstrate the odour prediction result of both models, as presented in Table 1 for the shallow neural network and Table 2 for the convolutional neural network. These tables are the two model results for the same dataset. Two performance evaluation methods are studied here, the cross entropy and the precision. The cross entropy loss is used to measure the model performance by comparing the predicted outputs with the target outputs. It is calculated through a log equation. For a classification model, cross entropy of 0 indicates perfect results. In the shallow neural network model, the cross entropy is 0.0269 for Table 1, and 0.0183 for Table 2 with the CNN model. Precision is defined as the accuracy of positive predictions and calculated on a per-class basis. The precision of urine odour prediction is 0.93 for the shallow neural network and 1.0 for the CNN, which is ideal. Therefore, while both have high overall accuracies, 92.56% for the shallow neural network and 97.22% for the CNN, the CNN provides better results not just in the urine odour detection, but also the predictions of other similar compounds. The shallow neural network has its merits in terms of easy implementation of a real-time classification device for incontinence individuals, and low memory requirement with relatively high accuracy. Overall, this work proves the feasibility of using these low-cost metal oxide sensors for urine prediction in an open environment that could contain interfering VOCs such as solvents.