Advanced decision support for complex clinical decisions

A Physician’s decision-making skills are directly related to the patient’s positive outcomes. Therefore, a wealth of medical knowledge and clinical experience are key assets for a physician to have. The goal here is to use historical clinical data and relationships processed by Artificial Intelligence (AI) techniques to aid physicians in their decision making process. Presenting this information in a Clinical Decision Support System (CDSS) is an effective means to consolidate decision results. The CDSS provides a large number of medical support functions to help clinicians make the most reasonable diagnosis and choose the best treatment measures. Initial results have shown great promise in accurately predicting Fibrosis Stage in Hepatitis patients. Utilizing this tool could mitigate the need for some liver biopsies in the more than 170 million Hepatitis patients worldwide. The prototype is extendable to accommodate additional techniques (for example genetic algorithms and logistics regression) and additional medical domain solutions (for example HIV/AIDS).


INTRODUCTION 1.CDSS Definition
In hospital information systems (HIS), there are typically two main systems: Hospital Management Information Systems (HMIS) and Clinical Information Systems (CIS) [1].HMIS support the hospital administration and transaction processing services while the CIS is used to support the clinical staff activities, to collect and dispose of clinical medical information, and to accumulate rich clinical knowledge.The CIS also provide clinical advice, support clinics, assistant clinical decision-making and to enhance staff efficiency.Clinical decision support systems (CDSS) are part of the CIS.It is an information system which uses expert systems and artificial intelligence (AI) technology to support clinical decision.It makes integrated diagnostic and medical advice bases on the collected patients' information, providing reference for the clinical medical officers.

Key Functions
Clinical decision support systems vary greatly in their complexity, function, and application.A Recent study [2] on health care information management four key functions of CDSS were outlined as follows: 1) Administrative: Supporting clinical coding and documentation, authorization of procedures, and referrals.
2) Managing clinical complexity and details: Keeping patients on research and chemotherapy protocols; tracking orders, referrals follow-up, and preventive care.
3) Cost control: Monitoring medication orders; avoiding duplicate or unnecessary tests.
4) Decision support: Supporting clinical diagnosis and treatment plan processes; and promoting use of best practices, condition-specific guidelines, and population-based management.
Our project will focus on item four, the decision support function and, in particular, utilization of historical laboratory data and outcome data processed through artificial intelligence tools.The combination of historical data and predictive tools provides valuable information in the hands of physicians as they develop a course of treatment for a patient.

BACKGROUND: AI TECHNIQUE IN CDSS
Decisions about medical treatment are best made by a trained and experienced physician.These decision makers can benefit from historical data and artificial intelligence tools.Computer scientists have dreamed of creating an "electronic brain" [3].Computer scientists and doctors alike have been captivated by the potential such a technology might have in medicine [4].With intelligent computers able to store and process vast stores of knowledge, the hope was that they would become perfect 'doctors in a box', assisting or surpassing clinicians Copyright © 2010 SciRes.

JBiSE
with tasks like diagnosis [3].Today the importance of diagnosis as a task requiring computer support in routine clinical situations receives much less emphasis.The strict focus on the medical setting has now broadened across the healthcare spectrum, and instead of AI Medical systems, it is more typical to describe them as CDSS [3].
In our project, we evaluated and compared two techniques that will be core forecasting tools in a CDSS, which are Data Mining and Neural Networks.The main purpose of doing data mining and knowledge discovery on the medical database is to predict disease and disease classification.Classification and prediction are two forms of data analysis which can be used to describe the model of the important data type or predict the future trends of the data [5].
Commonly used data mining algorithms are: association rules, decision trees, rough sets, statistical analysis, neural networks, support vector machines, fuzzy clustering, Case-Based Reasoning (CBR), Bayesian forecasting and visualization technology [6].The common methods used in auxiliary diagnosis of clinical disease are 1) Bayes discriminate analysis 2) artificial neural network 3) decision tree.
However within the context of the study, the focus will be concentrated on decision tree and neural network.

Decision Tree
The decision tree is a very efficient machine learning classification algorithm.It is the origin in the concept of learning systems CLS, and then progress to ID3 method.In the end, it evolved to c5.0 which can handle continuous attributes.Well-known decision tree methods are CART and Assistant [5].
Decision tree learning uses a decision tree as a predictive model which maps observations about an item to conclusions about the item's target value.In these tree structures, leaves represent classifications and branches represent conjunctions of features that lead to those classifications.One of the biggest advantages of the method is that the learning process does not require the user to understand a lot of background knowledge [6].
Nonetheless, data mining is a complex process to identify the useful information from large data sets.Although it is common to focus on the development, analysis and application of algorithms, the data selection and data pre-processing are the most timeconsuming activity in the entire data mining process, which affects the process and results [6].

Neural Network
Artificial Neural Networks have been proven to build efficient rule extraction/classification and forecasting applications.They provide a powerful non-linear machine learning techniques and are able to extract relevant features from large data sets.They are able to utilize and compare equally quantitative and qualitative data which is common in the clinical environments [7].Neural Networks can handle redundant features as weights are learned from the training data.
The primary disadvantage of neural networks is that they will always arrive at a solution for any data set.This means that the quality of the resulting model is highly dependent on the quality and breath of the training data.It is easy to over-train the model, so that it can only predict the training data set.This can be minimized by assuring that convergence stopping strategies are effective and that the training data set is representative of the solutions space.Also, the resulting neural network solution is simply a node structure with inter-node weights.This requires validation of the output by additional statistical methods (i.e.decision trees) that are more understandable by subject matter experts.

AI ASSISTED CLINICAL DECISION SUPPORT SYSTEM
This study will focus on the demonstration and incorporation of neural network and decision tree techniques into a Clinical Decision Support System.These two AI techniques were selected because of their complementary attributes.Neural networks provide little definition of their predictive result, while decision tree output provides clear connections to historical data.Both methods will provide different information to the physician.Other AI methods should be considered in future work.
The database we plan on using for this study was collected at Chiba University hospital in Japan, and is a Practice of Knowledge Discovery in Databases (PKDD) 2005 Discovery Challenge dataset [11].The data set contains patient data, laboratory data, and liver biopsies data on 771 hepatitis B and C patients.The goal will be to evaluate whether laboratory examinations can be used to estimate the stage of liver fibrosis.If this is possible, physicians may be able to use laboratory examinations as substitutes for biopsies and to aid in the treatment scenario.Liver biopsy is an invasive procedure and entails risk to patients.Decision tree and neural network methods based on a historical dataset will aid physicians in the development of treatment plans for Hepatitis patients.
The overall architectural representation of the system is shown in Figure 1.
The system is designed to be utilized by a physician to assist them in the development of a treatment plan for Hepatitis B and Hepatitis C patients.This system addresses an important question for the treatment -"Should a liver biopsy procedure be conducted?"This procedure provides important information on the advancement of the disease and fibrosis formation which are the key outcome measures.These fibrosis stage results determine the treatment protocol.This study attempts to predict fibrosis stage base on laboratory and patient data.The methodology and approach of processing data, applying AI techniques, and development of the resulting knowledge base can be utilized as a pattern for other medical treatment needs represented in a CDSS.

Data Processing and Cleaning
The sample hepatitis data set for our study is derived from the ECML/PKDD 2005 Discovery Challenge found at [11]: The hepatitis dataset contains the results of laboratory examinations taken on the patients of hepatitis B and C, who were admitted to Chiba University Hospital in Japan.Hepatitis A, B and C are virus infections that affect the liver of the patient.Hepatitis B and C chronically inflame the hepatocyte, whereas hepatitis A acutely inflames it.Hepatitis B and C are especially important because they have a potential risk of developing liver cirrhosis or hepatocarcinoma.An indicator that can be used to know the risk of cirrhosis or hepatocarcinoma is fibrosis of hepatocyte.For instance, liver cirrhosis is characterized as the terminal stage of liver fibrosis.
We utilized three tables for our study, as show in Tables 1-3, below: The MID field provided a common link between the three tables.Since the question we were addressing is to evaluate whether laboratory examinations can be used to estimate the stage of liver fibrosis our goal was to obtain one table that contained patient information, fibrosis stage and laboratory data.We chose to utilize in-hospital laboratory data, as it was more complete and we felt it would have better controls on quality and consistency.
Because of the large volume of data all files were translated into MS Access tables to process.The following steps were utilized: 1) Select all records from Table 2 -649 records.
2) Link to Table 1 based on Patient ID -649 records.
3) Link to the Laboratory Examination Table 3 based on Patient ID.Select records where laboratory examinations were available for at least 90% of our biopsy patients, resulting in 16 examination parameters.
4) Link to and Match appropriate lab examination re-sults (Table 3) with patient information and fibrosis stage information (the combined Tables 1 and 2).Select the laboratory examination that was earlier that the biopsy, but not more than 30 days earlier.This resulted in 425 records with complete data on the 16 laboratory examination results, patient information, and fibrosis stage biopsy results.

Methods and Analysis
The objective is to use the decision tree and neural networks to predict fibrosis stage from patient data and laboratory data.We have a data set (see data preparation) of 424 historical that we will use to train and validate a decision tree model and a neural network model.Table 4 is a map of the data we will be using.

Decision Tree Analysis
Data Mining is an analytic process designed to explore data (usually large amounts of data -typically business or market related) in search of consistent patterns and/or systematic relationships between variables, and then to validate the findings by applying the detected patterns to new subsets of data.Decision Tree's are one methodology utilized in the data mining set of tools.The ultimate goal of data mining is prediction -and predictive data mining is the most common type of data mining.
For the project, Weka 3.4 [9] which is a collection of machine learning algorithms for data mining tasks is used.The algorithms can either be applied directly to a dataset or called from your own code.Weka contains tools for data pre-processing, classification, regression, clustering, association rules, and visualization.It is also well-suited for developing new machine learning schemes.It also provides Decision Tree capability that we will use in this study.

Decision Tree Testing Procedure
Step 1: Data preprocessing The input data file needs to be organized in the form of ARFF in order to be processed in the Weka environment.In the file, all the values for the attributes needed to be filled in.If there is any missing value for an attribute, a "?" is used for substitution.
The Figure 2 shows the attributes and the patient's information whose Masked ID is 1.In the input file, even the data for the predicted attribute Biopsy Fibrosis needs to be completed for the training data set.For the analysis 392 patients' information was used to develop the decision tree module and predict the other 32 patients' fibrosis stage.
The remaining 32 patients' data was formatted into a test ARFF file.In this file the fibrosis is substituted with a "?". Figure 3, below shows the patient's information whose MID is 907.
Step 2: Build the Model and get Decision Tree We used the C4.5 algorithm to construct the decision tree.The root node is the Biopsy Fibrosis with five branches which present the 5 levels where the different laboratory values and patient characteristic are assigned one level at a time.Figure 4 shows a graphical presentation of the complete decision tree model of the data set.The Figure 5 shows a graphical presentation of the branch of the decision tree model where CHE <= 4.48.Step 3: Prediction The result of predicting the values with the constructed decision tree model is shown in Table 5.The table contains 32 patients' fibrosis stage values; the first column is the actual fibrosis stage, α, from the biopsy; the second is the predict values, β, by the decision tree model; and the third column shows the difference, γ = α -β.
The decision tree model results showed an accuracy of 37.5% (12/32) of correct fibroid prediction.Predicting fibrosis values within a range of +or -one of the actual fibroid stage showed an accuracy of 91% (29/32).

Suggestions to Improve Decision Tree
Accuracy 1) An increase in the number of training data examples will increase the correctness.Hence we can get a more powerful model, if we have a comprehensive training data set.
2) It is possible that obtaining additional laboratory attributes will increase the correctness.We were able to use only 16 laboratory attributes to predict the result.This was limited by the original data availability.We deleted some laboratory data since the incomplete data may also affect the test result.
3) The parameter setting for the decision tree algorithm utilized by Weka software is entered manually.We used the default value for the test because of limited time for the analysis.We may be able to improve results by testing additional parameters.

Neural Network Analysis
Artificial neural networks (ANNs) are systems that are constructed to use some organizational principles resembling those of the human brain.They are information processing systems that demonstrate the ability to learn, recall, and generalize from training patterns or data.ANNs are good at tasks such as pattern matching and classification, data clustering, and forecasting.
For the experiment we used a freeware tool called Neuro 3 [10]    Copyright © 2010 SciRes.

JBiSE
Step 2. Define the Neural Net architecture.The number of input nodes is defined by our 19 input (three patient data parameters and 16 laboratory parameters) rameters.And we have one output node (fibroid stage).The number of nodes in the hidden Layer is shown in the Figure 6.
Step 3. Define the run parameters.For this study we will only adjust the number of iterations that the error will be calculated and the weight adjusted.This error calculation and backpropagation will be executed 10,000 or 30,000 times.
Step Step 5. Utilize the trained neural network in Neuro 3 to predict the fibrosis stage for the prediction data set.Evaluate the goodness of fit by evaluating the percent of predictions that were correct, and that were within plus or minus 1, 2, or 3.
Step 6. Adjust the parameters in Step 1 through Step 5 until the optimal fit is achieved.
Step 7. Utilize the final selected model and train this model with all 424 historical data elements.Identify any deficiencies in the process or items that would be helpful for future work.

Testing and Validation Results
Eleven neural network models were generated.The best  fitting model is Run Number 8, a 19-4-1 architecture trained with 10,000 iterations.This model provides fairly good predictive capabilities with 56% of predictions being correct (γ= α -β = 0) and 90% being within +/one fibroid stage (γ= α -β = ± 1).The run statistics are show below in Table 6.
The process of testing and validation of results revealed the following future considerations: First, the increase in the amount of training data would greatly improve the predictive capability of the tool.This can be observed comparing Run Number 1 to Run Number 6.The application should allow for inclusion of additional data.
Second, additional data elements may increase the predictive capabilities of the tool.We only had sufficient data coverage for 16 laboratory data elements.Additional laboratory data elements should be acquired and evaluated utilizing these techniques.
Third, the analysis tools and selection of neural network architecture and run parameters was selected manually.This was a limitation of time available for this study.Automated analysis techniques and hybrid techniques combining neural network and genetic algorithms should be considered.An illustration of the sensitivity of predictive capabilities to hidden layer nodes shows in the Figure 8 below.The vertical axis shows the percent of patients that the neural network predicted the correct fibrosis stage.The horizontal axis shows the number of hidden nodes specified in the neural network model, indicating that the accuracy of prediction is very sensitive to the number of hidden nodes in the model.

CONCLUSIONS
We have completed the development of a prototype that utilizes publically available patient data to address a single clinical decision-the prediction of fibrosis stage shown in Figure 9.The near-term customers for our project prototype are clinicians treating Hepatitis B and C patients.
Currently the key diagnostic tool in assessing the degree of liver disease in these patients is a liver biopsy.This procedure is invasive and requires the physician to

Figure 1 .
Figure 1.AI assisted clinical decision support system.

Figure 5 .
Figure 5. Section of the decision tree.
4. Run the neural network model on the training data set.The Neuro 3 tool provides a R-squared value and Sum of Errors for the training data set that provides some indication of the goodness of the model.A example of the Neuro 3 screen is shown in Figure 7.

Figure 6 .
Figure 6.Architecture of a BPN neural network.

Figure 8 .
Figure 8. Correctness as a function of the number of hidden nodes.

Table 4 .
Data utilized for neural network.

Table 5 .
Decision tree prediction quality.
which uses the back propagation neural network (BPN).The term backpropagation refers to the training method by which the weights of the network connection are adjusted.The calculations procedure is feedforward, from input layer through hidden layers to output layer.During training, the calculated outputs are compared with the desired values, and then the errors are backpropagated to correct all weight factors.