World's first medical networking and resource portal

Category : All
Medical Articles
Artificial Intelligence to predict BP using Leucocytapheresis
Artificial Intelligence to predict BP using Leucocytapheresis

BP prediction during Leukocytapheresis (LCAP) using Artificial Intelligence



With rapid advances in information technology, computer-aided medical treatment is becoming quite common in hospitals. Several evolved treatments rely on artificial intelligence for prediction of various parameters that affect the blood purification [3]-[5]. During blood purification treatment, it is extremely necessary to monitor circulating blood volume of patient. Till now, most of the hospitals rely on the medical workers to use their own judgment in order to make necessary changes that affect the stability of the patientís blood pressure. However, this approach is very subjective and could lead to undesirable fluctuation in blood pressure. Hence, a computer-aided method can provide steady blood purification by predicting blood pressure values. This could be an immense valued data to medical workers in providing accurate treatment. Such a method using artificial intelligence is proposed here for evaluation.

Artificial Neural Networks (ANNs) is gaining popularity in prediction methods. These are arrays of simultaneous equations that iteratively examine data sets according to learning rules.[Pandya and Macy, 1995] Delta rule is the most commonly and extensively studied prediction method that performs gradient descent optimization and is thus closely related to standard regression models. ANN using the delta rule has been successfully applied in predicting outcome in a variety of complex biomedical problems. ANNs have been widely used in solving real biomedical problems because of their ability to learn the dynamics of complex systems and to identify multidimensional relationships among multiple variables from available input/output samples.

References: IUGR: [7] Gurigen F et al., IEEE Eng. Med. Biol. 1997; 16:55-58, studied ultrasonographic examinations using neural networks (NN) in the detection of intrauterine growth retardation (IUGR). They concluded NN is a very helpful tool for correlating many variables. ECG: [8] Maglaveras N et al., IEEE Trans Biomed Eng. 1998; 45:805-813, worked on adaptive back propagation neural network for real-time ischemia episodes detection, development and performance analysis using the European St-T database.

Coronary artery disease: [9] Shen Z et al., Comp Cardiol 1993; 20:221-224, A neural network approach in the detection of coronary artery disease. Liver transplantation outcomes prediction: [10] Doyle HR et al., Ann Surg 1994; 219:408-415, Predicting out-comes after liver transplantation: A connectionist approach Electronic noise: [11] Kermani BG et al., IEEE Trans Biomed Eng 1999; 46:429-439, Using neural networks and genetic algorithms to enhance performance in electronic noise..

A Neural Network model can facilitate better treatment by predicting the change in circulation of blood volume. ************ The neural network is trained to predict the Ht value after various intervals of time. If the neural network predicts a drop in Ht value, it would indicate that the blood pressure is likely to drop and the medical worker can make the necessary adjustments.

Ulcerative colitis (UC), as well as Crohnís disease (CD), is one of the prototype nonspecific inflammatory bowel diseases (IBDs) of unknown cause. Drugs such as salazosulfapyridine, 5-aminosalicylic, and immunosuppressants have been used for the treatment of UC. These drugs attenuate the inflammation in the colonic mucosa by their anti-inflammation and immunosuppressive actions, and causes disease remission. However, some patients with UC are even refractory to the strong anti-inflammatory and immunosuppressive actions of steroids. Surgical treatment often has been considered as the ultimate treatment modality for such patients.

Extracorporeal circulation treatment methods have shown to be highly effective for the treatment of UC patients refractory to steroids. One such method is called as granulocyte and monocyte adsorption apheresis (GMA), in which mainly granulocytes and monocytes are removed from the blood. Another method is leukocytapheresis (LCAP), in which granulocytes, monocytes, as well as lymphocytes, are removed from the blood. These treatment modalities have been reported to yield a high therapeutic efficacy in many patients of UC, including those who are refractory to steroids.

Leukocyctapheresis (LCAP) is a blood purification treatment for ulcerative colitis (UC) [1]. LCAP is known to have a low incidence of side effects. During treatment the blood pressure of patient starts decreasing. This phenomenon is caused by change in circulation of blood volume. The continuous reduction of the hematocrit value (Ht) during blood purification process lowers blood pressure. Hence it is a critical factor and is considered to be a major determinant of infusion dose and vice-versa, where artificial neural network (ANN) has an edge over individual judgment.

LCAP is carried out using a column (Cellsorba E) filled with a non-woven fabric made up of polyester fibers. The fabric had a dual structure; an inner layer composed of superfine fibers 0.8-2.8 in diameter, and an outer layer composed of fibers 10-40 in diameter. The blood is filtrated from the outside into the inside of the non-woven fabric wound into a cylindrical shape in the column, and leukocyte components are removed. The blood, with leukocyte removed, is guided out from the column and heated, and the returned to the corresponding vein of the patientís other arm or leg of the patient. The blood flow rate is set at 30-50, and 2-3 L of blood is treated in each session of LCAP. The treatment is carried out for one hour per session once in a week, for 10 wks.

In this paper an ANN model is proposed as a predictor of the Ht values after every 1, 3 and 5 minutes. Well-known Multi Layer Perceptron (MLP) neural network (NN) is used here for data evaluation. (Fig.1). MLP is an appropriate instrument to deal with this kind of problem because it can handle the intrinsic nonlinearities involved in these types of biological systems. This in process identifies multidimensional relationships and learns the input/output characteristics from input/output samples. Different input parameter combinations have been tried in order to find the most effective model for prediction.

Currently, a serial Hematocrit Monitor manufactured by CRIT-LINE Monitor (CLM), Hema Metrics Inc Salt Lake City Boston U.S.A is in use due to its particular capabilities. It facilitates the noninvasive monitoring of the Ht value, rate of change in the circulating blood volume (percent blood volume change,BV %), and venous blood oxygen saturation. This is done after every 20-second intervals by measuring the absorption rate of scattered infrared rays in a chamber set within the blood circuit of the blood purification devise. In the Tokushima University Hospital, CLM is used for obtaining Ht values during blood purification. Table1 shows the index of CLM [2].

Berns et al (1999) [13] used interdialytic ambulatory blood pressure (ABP) monitoring to study the effects of partially corrected anemia versus normal hematocrit (hct) on BP in hemodialysis patients. They report that the mean daytime and night time BPs were not different from each other at two, four, and eight months in anemic group or at any time in normal group, and in both groups, most patients had little diurnal change in BP. However this study focuses on BP normal daily activities, i.e. ABP, while here the focus is on BP of an individual as it relates to changes in Ht values during blood transfusion.

L LaRue et al (1987) [14] found no significant difference in BP between hematocrit and stroke subtypes in normotensive individuals either in low (less than or equal to 30, 30-36%) or high (greater than or equal to 47%) hematocrit groups.

Martin S et al (1971) [15] studied to assess the importance of an elevated cardiac output in the generation of the hypertension associated with chronic renal failure. It was concluded that the elevation of cardiac index in uremic patients is secondary to anemia and is reversible when the hematocrit is raised over 30%. The high cardiac index is not responsible for hypertension because restoration of cardiac index to normal by transfusion raises blood pressure rather than lowers it.

Giovanni Bertinieri et al (1998) [16] studied reduction in blood viscosity without changing blood volume causes a significant fall in both clinic and 24-hour ambulatory BPs. This is particularly true when, as can often happen, blood pressure is elevated. This emphasizes the importance this variable may have in the determination of blood pressure and the potential therapeutic value of its correction when altered.

Judith Martini et al (2005) [17] studied Hematocrit (Hct) relation between Hct and blood pressure and these findings suggest that increasing Hct increases blood viscosity, shear stress, and NO production, leading to vasodilatation and mild hypotension. Larger increases of Hct (>19% of baseline) led blood viscosity to increase >50%, overwhelming the NO effect through a significant viscosity-dependent increase in vascular resistance, causing MAP to rise above baseline values.

LACP Treatment Methodology & Application of Neural Network

The multi-layer perceptron designed for our application had linear function with regard to the input and output layer. The hidden layer neurons had a non-linear function, which was chosen as tanh in this case.

The prediction accuracy depends on the acted structure of the neural network and was examined for total 120 different types of them. The networkís performance also depends on the initial weight values. As a result, for each data set, the network was trained with 20 different initial weight values.


The Neural Network Prediction Method

In each data set, Ht values of first 15 minutes at the beginning of LCAP procedure were not used. This was based on the fact that infusion of normal saline solution lowers the accuracy of measured Ht values during the initial period.

In this study, biological time series signals after 15 minutes of starting LCAP are represented as:

&nb sp; (1)

Where is sampled data at sampling time and is the last Ht value. The Ht data were divided into the consecutive pattern groups for improving the accuracy of prediction. A pattern group is defined as:

&nb sp; (2)

Where is index of group. is sampling time, and is training period. In this study, 60 Ht values were used for the training period. Pattern groups were formed by using Ht values in the following manner:

&nb sp; (3)

In this study, the neural network model is moving average type. The input/output relation of the neural network in the pattern group is as follows:

       Input &nb sp; | output

&nb sp; (4)

Where shows the number of input units and shows the prediction time. For example, the sampling interval is 20 seconds, hence means that neural network model will predict Ht value after each 1 minute.

When the number of units in input layer is, the neural network is trained to output based on input. Therefore in pattern group the neural network is trained using vectors. The connection weights were used as initial weights. were small random values. The back propagation algorithm was used as the supervised leaning algorithm for the Multi Layer Perceptron (MLP). When the following condition (Eg.5) was fulfilled, the training was ended.

and &nb sp; (5)

Where is for ideal condition. However it can be set to a value lower than one for tolerance.

Where was indicates the rms error during the training period which is given by

&nb sp; (6)

and is the number of sample in the training set.

Here is the weight matrix for pattern group at iteration during the training. In order to develop a moving average type NN model first the network was trained with all the pattern vectors in pattern group . When the network converges, the weights are obtained. Then the training is carried out using pattern vectors in Pattern Group and is used as initial weights. The sequence is repeated until the training is completed using the group.

The prediction is an output of the neural network model corresponding of input vector.

Error between the measured Ht values and the predicted values was defined as:

&nb sp; (7)

Where is number of pattern of groups.


The prediction accuracy was examined using 120 kinds of NN structures. At first the number of units in the input and hidden layer were varied from 1 to 15. However, the number of units in the input layer always doesnít to be more than the number of units in the hidden layer in order to avoid over training. Hence, the neural network with the smallest rms error in 120 kinds of structures was chosen as the best NN for the prediction. The NNís output usually depends on the initial values of connection weights. Therefore initial weights were changed 20 times for each data. The average and standard deviation of rms error in 1, 3, and 5 minutes later were shown in Fig.3, Fig.4, and Fig.5. NN# shows the various NN structures. From Fig.3, Fig.4 and Fig.5, the most suitable NN structure has 1 unit in the input layer and 1 unit in the hidden layer 1 minute later. The most suitable NN structure has 1 unit in the input layer and 1 unit in the hidden layer 3 minutes later. The most suitable NN structure has 2 units in the input layer and 1 unit in the hidden layer 5 minutes later. The results of the prediction using these NN structures at the patient A shown in Fig.6, Fig.7 and Fig.8. Table3 shows the average and deviation of rms error between the measured Ht value and the predicted Ht value about each patients. From Table3, all results were small rms error between the measured Ht data and the predicted Ht data using NN.

Discussion and Conclusion

In this study, it was examined that the change of Ht value 1,3, and 5 minutes later can be predicted using NN. Ht values were predicted using 120 kinds of NN structures, but the best NN structure for prediction was the simple NN structure. The results were achieved since the change of Ht value in LCAP was small. All predictions of Ht values using NN were small rms error in Table3. It is dangerous if there is an error of 1% or more in Ht value in blood purification treatment. Therefore, it seems to be acceptable prediction of Ht values in all cases. This results shows the neural network predicts the Ht successfully.

Dr Tejinder Mohan Aggarwal

FORMER: *Research Associate,
CS & E, Florida Atlantic University (FAU), Boca Raton Fl 33431 USA

Category (Brain & Nerves)  |   Views (11982)  |  User Rating
Rate It

Browse Archive