Modified Neocognitron Network for Medical Signal Classification

A modified neocognitron neural network suitable for medical signal classification is presented. The network's functionality is demonstrated on an application involving the classification of breathing signals measured on patients recovering from surgery. The performance of the system was found to be equivalent, and in some cases, better than a standard technique used for comparison. The main advantage of this system is that it allows for a detailed analysis into the decisions made by the system. Additionally, its internal property enables the rejection of inputs which are found to be too different to the learned ones, the thresholds for rejection can be set by tuning the structure of the network and selectivity of the cells in the network. This makes it possible to detect artefacts or identify new classes that might exist, in addition to providing the appropriate classification in the final step of the processing chain. Figure 1: Structure of the Neocognitron Figure 2: Connection between two planes 2. The size of this area is typically 3x3, 5x5 or 7x7 neurons. This square area shifts in parallel with the position of the target neuron, thus preserving topographical information between layers. All the neurons in a plane share the same set of weights. Sub-layer S contains the so-called feature extracting cells. All the cells in a single plane are specially adapted to detect the presence of a particular feature in their input area. An important parameter related to S cells is their selectivity. It expresses the level of tolerance of the cells. The higher the selectivity, the more precise the match between the inputs of the cell and the trained feature that is required for a neuron to fire. Examples of features for a four-layer neocognitron designed for handwritten numbers recognition [3] are shown in figure 3. The features are shown as they would appear in the input layer. These features were created by the author of [3]. The connecting rule of sub-layers C is the same, but they play a different role. They “blur“ the information received from the feature extracting cells to make the network more invariant to distortion of the input examples. Their weights are fixed and set accordingly. Globally, the network is hierarchical. The first layer extracts very simple features like small lines of different orientations as shown in figure 3. The next layers extract features of increasing complexity. In the neocognitron as described in [3], the features for the different layers are pre-defined, and they are not created or modified by training, so training examples are not required for the network. The size of planes in the last C sub-layer is only 1x1. There are 10 planes in this sub-layer, one plane per class. Here the plane with the highest output represents the classification output of the neocognitron. Outputs of all different planes and cells in the original neocognitron after propagating the number “4“ are shown in figure 4. More details about the neocognitron can be found in [3] and [4]. MEDICAL SIGNALS AND THEIR PREPROCESSING In a study aimed at monitoring postoperative patients, the relationships between respiratory pressures and dimensions were visually analysed [1] to show that it might be possible to investigate how analgesia, airway obstruction and hypoxia (lack of oxygen) are correlated. Five physiological signals: oesophageal (Poes) and gastric (Pgas) pressures, chest (CHST) and abdomen (ABDO) dimensions and nasal (Nasal) air flow were recorded and their phase relationships were studied. The following information in this section, also described in [1] and [2], is repeated here for clarity and ease of explanation. Asynchrony of ribcage and abdominal movement was found to be related to airway obstruction [1]. In normal breathing, gastric pressure and abdominal dimension start to rise simultaneously with the dimension of the chest at the onset of inspiration. This was categorised as Class A. Two additional mechanisms, which can be interpreted as abnormal, were also observed. The first involved a gastric pressure paradox, where gastric pressure decreases at the onset of inspiration marked by an increase of the chest dimension and a phase lag in abdominal dimension, this was categorised as Class B. In the second, the gastric pressure and the abdominal dimension decreases at the onset of inspiration, this was categorised as class C. In previous work [2] self organising maps (SOMs) were used to learn the graphical representations of the relationship between the signals defining the different categories of the breathing mechanism. It was shown that using a combination of signal pairs, the SOMs self organised into groups of nodes responding to similar categories as those introduced by Nimmo and Drummond [1]. Experiments were conducted to identify the pairs of signals that gave the best separation of categories. The data used included signals from patients aged 40 or more, who had undergone major abdominal surgery involving an incision at least partly above the umbilicus. The exclusion criteria were patients with severe cardiac, respiratory or renal disease or an abnormal body weight. The signals were measured during the first night after surgery under observation at the Edinburgh Royal Infirmary. The average duration of the recording of each patient was about 5.5 hours. Data from a total of eight patients was used. This was the same data set as used in the previous studies [1], [2]. Chest and abdominal movement signals were measured using inductance bands placed around the chest and abdomen. Oesophageal and gastric pressure were measured using a modified nasogastric tube with an integral oesophageal balloon and a gastric balloon attached to the tube tip. Table 1: Three classes of breaths Pgas ABDO Class Normal normal A