A decision support system to determine optimal ventilator settings

The use of decision support systems [1] in health provides improvement in terms of the quality of health and care, early diagnosis of diseases, prevention of person related errors, lowering costs and providing the patients with the optimum treatment.

Today, most physicians prefer the decision support systems in the interest of the effective operation and early determination of the most appropriate option in order to avoid the growing problem related to the management of medical information. The decision support systems are the most ideal aids in either detection or treatment of a disease, or determination of the most appropriate drug.

A basic requirement of our survival is to take a sufficient amount of air and to transfer it to all our cells. While heart and blood cells are responsible for distribution, our lungs perform the exchange of oxygen (O ₂) and carbon dioxide (C O ₂) by holding a medietary position between the atmosphere and the internal organs. The delivery of oxygen and removal of carbon dioxide, which are the two most important components of the inspiration - expiration process is, called “breathing” [2]. The respiratory function is autonomously carried out in healthy living beings and is artificially carried out in those with lung disease and respiratory failure. The device that is used in this process is called as the ventilator device [3] and the process is called ventilation.

The physiological parameters determined in accordance with the findings of the patient who is connected to the ventilator device change simultaneously depending on the instantaneous status of the patient after the device has been inserted. The most critical point here is to determine the device parameters in the desired accuracy. Since the task of specifying the parameters of ventilation equipment is carried out entirely by a physician, the physician’s knowledge and experience in the selection of these settings have a direct effect on the accuracy of his/her decisions. The process of determination of the device parameters with a decision support system to be developed for this decision process can be completed with minimum error by predicating the instantaneous status of the patient. By this way, the decision making processes of physicians can be implemented in a faster and a more accurate manner.

This study describes a decision support system designed to automatically determine the ventilator settings in consideration of the ventilation data obtained from two different hospitals. For the model system in our study, COPD (Chronic Obstructive Pulmonary Disease) and CVD (Cerebrovascular Disease) have been used as respiratory diseases. COPD [4] is a disease that causes the obstruction of the bronchi on the long term and that has no treatment. ARDS [5] is the acute respiratory distress syndrome caused by the increase in the alveolar capillary permeability. CVD [6] is a group of disorders with symptoms related to the damaged brain region associated with the blockage or bleeding of the arteries supplying the brain. Although CVD is not a respiratory disease, almost all CVD patients are treated with ventilation devices.

According to the Turkish Statistical Institute’s report of causes of death statistics [7], 16.311 patients died of disease caused by COPD (J40-J44) and 30.103 patients died of disease caused by CVD (I60-I69) in 2009. In total mortality of 2009, it corresponds to 16,5%.

For this reason, the management of mechanical ventilation devices used in intensive care units is a critical process. Our goal is to minimize errors in this process and prevent deaths caused by incorrect configuration of ventilation devices. This system is designed for the facilities without having lung mechanics like cottage hospitals in country regions. Metropolitan cities have fully equipped hospitals with experienced medical personnel, but on the other hand less experienced doctors can benefit from this system. Our proposed method is facilitated upon making the medical data of patients meaningful with help of an Artificial Neural Network (ANN) [8], and the determination of the preferred treatment factors.

Below, further studies conducted so far with regards to the subject of our study have been summarized.

In the study of Adeney, Ennett, Frize and Korenberg [9], this methodology is used and a model using two-stage ANN is presented. It is observed from the test results that a high success rate is provided with regards to the estimation of the parameters.

A similar study has simultaneously performed the inspiration – expiration of the intensive care patients by a ventilator device [10]. The model working in accordance with the data such as the patient’s breathing frequency, the density of oxygen in their blood and the flow wavelength of the blood has estimated the ideal working time.

Borrello’s study [11] is a detailed modeling of a prototype which is thought as a replacement for the standard ventilator devices and which provides assistance for the operator with regards to the parameters to be selected through its own processor and sensor.

In the study of Tehrani [12], a closed-loop decision support system that foresees the leaving time of the patient from the ventilator depending on values P C O ₂ (Partial Pressure of Carbon Dioxide), P S O ₂ (Partial Pressure of Sulfur Dioxide), F i O ₂ (Fraction of Inspired Oxygen), PEEP (Positive End-Expiratory Pressure) parameters was proposed.

In another study [13], simulations have been implemented in the MATLAB environment by using the hybrid algorithm for the ventilators used in intensive care units. With the designed model, the settings regarding the parameters of the blood gasses like F i O ₂, PEEP, Pinsp, Vrate are recommended to the physician by the system.

Another computer aided ventilator prototype study [14] was developed by Ahmedi and Bates. Prototype ventilator opens and closes the valve by the help of the computer program, which sustains the process of inspiration and expiration.

Zhu [15] and Allerød [16] proposed ventilation settings with a decision support system according to the patient’s physiological characteristics of the patient to be undergone ventilation, by using ANNs. While Zhu has used the data from 28 ARDS patients in his study; Allerød used the data from 20 CABG patients. In both studies, it is shown that patient characteristics are important in the selection of the ventilator settings.

As a result, it is seen in the research articles [17] related to the methodology of the ventilator devices controlled by the smart control techniques that the influence of the decision support systems on the decisions of the physician have been quite successful.

Er, Yumusak and Temurtas [18] presented a comparative chest diagnosis; for chronic obstructive pulmonary, pneumonia, asthma, tuberculosis and lung cancer diseases which was realized by using multilayer, probabilistic, learning vector optimization, and generalized regression.

Gil et al. [19] used ANN models as tools for support in the medical diagnosis of urological dysfunctions. They developed two types of unsupervised and one supervised neural network to distinguish and classify between ill and healthy patients.

Ushida et al. [20] used fuzzy neural network analysis (FNN) of health check-up data to provide a personalized novel diagnostic and therapeutic method involving the y-GTP level and the WBC count. They performed a logistic regression analysis, including adjustment to ensure FNN analysis was statistically reliable.

Heckerling et al. [21] used ANN coupled with genetic algorithms to evolve combinations of clinical variables optimized for predicting urinary tract infection for women. Their method revealed that parsimonious variable sets accurate for predicting urinary tract infection, and novel relationships between symptoms, urinalysis findings, and infection.

The Faculty of Medicine, the Electronic and Computing Engineering Department of the Federal University of Rio de Janeiro (UFRJ) and the Electrical Engineering Department of the Federal Center of Technological Education (CEFET-RJ) worked on a collaborative research project to develop a decision support system for smear negative pulmonary tuberculosis (SNPT) [22]. This project aims to develop, through a multi-disciplinary, multi-institutional, innovative and cost-effective approach, new paradigms to prevent the disease progression and support the rapid evaluation of new therapies.

Moein, Monadjemia and Moallem [23], investigated typical disease diagnoses using ANNs. After selecting some symptoms of eight different diseases, MLP neural network was used with a fuzzy approach to get more accurate results.

Kamruzzaman and Islam [24], proposed an algorithm, called rule extraction from ANNs (REANN), to extract rules from trained ANNs for medical diagnosis conditions like breast cancer, diabetes and lenses problem.

Lisboa and Taktak [25] conducted to assess the benefit of ANNs as decision making tools in the field of cancer. Their study describes the work being done in this area over the last decade.

Clinical applications of ANNs provide benefits in many areas, and there are still many more perspectives to be examined like ethics and clinical prospects [26].

In the Methods Section, the methodology and design elements of the decision support system are detailed. Experimental results and conclusion are given in Section “Results and discussion” and Section “Conclusions”, respectively.

Prior to the second step of the system which is the estimation of the ventilator settings, the probabilities of the diseases were predicted and presented to the physician, and upon the verification of relevant disease or determination of a new disease by the physician, they were sent as parameters for the predicted model of ventilator settings.

Firstly, Levenberg-Marquardt’s backpropagation learning algorithm has been tested by using different number of hidden layers in order to train the regression model through which F i O ₂, frequency and tidal volume parameters were proposed. The prediction rates obtained by using the Levenberg-Marquardt algorithm were calculated as 86,72% for F i O ₂, 83,66% for frequency and 87,17% for the tidal volume, when 30 hidden layers were used for the highest prediction percentage. Operation of the system from 15 hidden layers to 100 hidden layers without decreasing the accuracy below the average of 80,63% was completed in approximately 1.5 seconds. The graphic showing test results of the Levenberg-Marquardt algorithm is given in Figure 4.

Following the Levenberg-Marquardt algorithm, another backpropagation learning algorithm used in test was the Bayesian Regulation. In all the preferred hidden layer parameters used during the tests, the average success was never below 78,91%. The most successful test results were obtained by using 50 hidden layers following a process period of 2 minutes which were recorded as 91,43% for F i O ₂, 89,6% for the frequency and 90,61% for the tidal volume. When all the tests are examined in general, it is seen that the hidden layer parameter in the Bayesian Regulation backpropagation learning algorithm directly effects the time of operation. The time of operation which lasted for 2,7 seconds when 5 hidden layers were used increased to 637 seconds when 100 hidden layers were used. The success rates of the tests where the Bayesian Regulation algorithm was used are given in Figure 5.

Third backpropagation learning algorithm used in the tests is the One step secant. The operation time in the tests that used hidden layers from 5 to 100 as parameters has been completed within a range of 15 – 20 sec. The general average success rate was calculated as 80,99%, even for the test where 5 intermediate layers were used. In the test where 50 hidden layers were used and which had the highest success rate, the values for F i O ₂, frequency and tidal volume were 90,81%, 59,58% and 91,59%, respectively. A graphic showing the test result of One step secant algorithm is shown in Figure 6.

The test results where the BFGS quasi-Newton backpropagation learning algorithm was used gave the worst estimation percentages. Even in the best conditions, the success rates for F i O ₂, frequency and tidal volume were 83,56%, 77,72% and 78,5%, respectively. These percentages were obtained from ANN using 20 different hidden layers. Test results of the BFGS quasi-Newton algorithm are shown in Figure 7.

Two supervised training algorithms have been used as alternative to backpropagation learning algorithms we used for training ANN. The first algorithm is the Cyclical order weight/bias algorithm. No estimation was provided where higher numbers of hidden layers were used. Even in cases where a comparatively low number of hidden layers were used, the lowest success rate was obtained among all the methods. In ANN configuration with an operation time of 126 sec where 5 hidden layers were used, the estimation success rates for F i O ₂, frequency and tidal volume were 67,41%, 73,46% and 74,31%, respectively. The graphic regarding the tests conducted with all the Cyclical order weight/bias algorithms is shown in Figure 8.

Recent tests made in the regression model have been realized by using the Sequential order weight/bias supervised learning algorithm. It was a general observation that this algorithm was capable of accomplishing the operation at high speeds and in consistency with the highest number of hidden layers. This provided the highest success rate among all the tests using other algorithms. As per the configuration where 300 intermediate layers were preferred, the prediction success rates were 100% for F i O ₂, 99,81% for frequency and 99,54% for tidal volume, after an operation time of only 6 seconds. Tests made by using this algorithm are shown in Figure 9.

In the tests for the classification of pressure support and volume support parameters, the highest success rates of 94,4% and 92,2% were obtained by the Bayesian Regulation backpropagation learning algorithm and the Cyclical order weight/bias supervised learning algorithm, respectively. When the operation time is considered as an important parameter, the Bayesian Regulation algorithm should be preferred in the first place due to having an operation time of 1.1 seconds. Values for the tests made for the classification model have been presented in Figure 10.

As seen from the examples provided in Table 3, when the results produced by the decision support system we proposed are compared to the ventilator settings actually used for obtaining data; it is shown that the success rates in the prediction of frequency parameter, tidal volume parameters, F i O ₂ parameters and pressure support/volume support parameter are 99,81%, 99,54%, 100% and 94,40%, respectively.

Furthermore, proposed models field tests conducted on 14 patients of 2 different doctors show that, the system trained with 158 patient data recommended the ventilator settings with 95,3% accuracy. This percentage has remained low compared to the simulation results, but this ratio provides a satisfactory grade for the use of our proposed system in real life. It is clear that, gaining more data to train the system increase this success ratio.

This article has aimed at making the parameter selection process in the ventilator treatment period of the respiratory tract patients independent from the physician, via the decision support system proposed. As seen from the experimental test results, the system can propose the frequency, tidal volume, F i O ₂ and pressure support/volume support parameters by an accuracy rate of 98,44% on simulations and an accuracy rate of 95,3% on field tests. The incorrect suggestions has small deviance to optimum values, which are harmless to patients in case of application.

The training of the ANN in the system has been realized with the physiological data from 158 patients. The rate of accuracy in the system’s prediction capability will increase with the use of data from higher number of patients. Thus, non-physician operators will also be able to use the system towards the determination of the ventilator settings in case of emergencies.

Improvement of the project as an automation system can be provided by the transmission of the physiological values of the patients to our decision support system from the hospital database and the direct installment of the treatment parameters preferred by the system in the ventilator device, without a mediator. For further studies we are going to identify the parameters which have less impact on results. Elimination of these parameters will speed-up the execution periods.