Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation. By: Jorge Garza-Ulloa Electrical & Computer Engineering Doctoral Program The University of Texas at El Paso [email protected] Abstract: Using Artificial Intelligence tools to predict three Stroke Variables: Surgery needed, Rehabilitation and Days of rehabilitation, with this information we will have a reference point and our general goal on future research is to a follow-up of the subjects to integrate them faster to their normal life to the patients and lower cost related for expenses of the illness”. Neural Network (NN) algorithms are proposed to develop a Database Acknowledgment (DBK) to predict the three parameters mentioned: Surgery, Treatment and Days of rehabilitation. Goal is to find an optimal Neural network configuration using with the actual information available using three different software available: one manual (with no automatic stepwise functions, limited diagnostic), another semi-automatic (allows stepwise function and good diagnostic) and one Neuro-Intelligence (use Genetic Algorithm to find the best NN configuration and an excellent diagnostic tools) our proposed solution must be minimum possible prediction error. Based on the 14 Stroke input variables and the 3 output target Stroke values, this paper suggest that the forecasting of: Surgery, Rehab and Days of Rehabilitation it is possible using Neural network tools. Introduction: “Every year, more than 750,000 people suffer a stroke, a “brain attack”. Stroke occurs when blood flow to an area of the brain is stopped. As a result, lifesupporting supplies of blood and oxygen are cut off from the brain. Brain damage due to a stroke can affect important areas that control everything we do – including how we move different parts of our body. ”The recovery of Stroke Patient’s is regularly very slow and expensive; any solution to reduce this time and/or prevent this illness will be very usefully for Human kindness”. Page 1 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Neural Network Neural networks are composed of simple elements operating in parallel. These elements are inspired by biological nervous systems. As in nature, the network function is determined largely by the connections between elements. You can train a neural network to perform a particular function by adjusting the values of the connections (weights) between elements. Commonly neural networks are adjusted, or trained, so that a particular input leads to a specific target output. An elementary neuron with R inputs, where each input is weighted with an appropriate w. The sum of the weighted inputs and the bias forms the input to the transfer function f. Neurons can use any differentiable transfer function f to generate their output. And get network with the n Input, m Hidden Layers and k Output Target as shown on the next figure: Page 2 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa We propose a supervised network that is trained to produce outputs in response to sample inputs, making them particularly well-suited to modeling and controlling dynamic systems, and predicting future events. Dataset for the proposed neural network: Use dataset available (Collected data) _ The Dataset to use is Stroke_full_database with 1048 samples: Many Eyes from An experiment brought to you by IBM Research and the IBM Cognos software [1] : From this dataset we choose the 14 variables forming 4 different groups: 1) Stroke Independent with three variables : AgeCat ( 4 values form 45 to 86 years old), Gender and Active ( still working or doing daily activities) see table 1 2) Stroke Risk Factors with 8 variables: Obesity, Diabetes, Blood Pressure, Atrial Fibrillation, Smoker, Cholesterol, Angina Chest Pain and Transient Iscg ( ministrokes). 3) Count of actual Strokes of the subjects with three variables with values from [0 to 1] : First Stroke, Second Stroke and Third Stroke 4) Stroke target variables , the three that we want to predict: Surgery, Rehabilitation and Length of Days Rehabilitation We summarized these four groups of variables and their ranges on figure 1. Page 3 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Figure 1 Table 1 4 Groups variables for the proposed NN To analyze this dataset we use WEKA Open Source tool for Data mining [3] these information is summarized on the next four tables: Table 1, 2 and 3 shown the Input variables and table 4 the Target Variables Variable AgeCat (Subject Age) Sub-Groups & Ranges [45,55]=.25 , [56, 65]=.50 , [65,75]=.75 , [76,86]=1 Gender Female=0, Male=1 Active No=0 Yes=1 Actual Data on Dataset Table 1 Dataset Pre-analysis: Input Stroke Independent Variables Page 4 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Variable Overweight ( based on normal weight) Diabetes BloodPressure Sub-Groups & Ranges No=0, Yes=1 Actual Data on Dataset No=0, Yes=1 Normal=0, normal high=.5 high=1 Smoke No=0, Yes=1 (AF) Heart palpitations No=0, Yes=1 Cholesterol No=0, Yes=1 Angina (Chest Pain) No=0, Yes=1 Mini-Stroke (TIA) No=0, Yes=1 Table 2 Dataset Pre-analysis: Input Stroke Risk Factor Variables Page 5 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Variable Stroke1 Sub-Groups & Ranges [Very Low, Low]=.25 [Low High, Medium Low]=.50 [Medium, Medium High]=.75 , [High, Very High]=1 Stroke2 [Very Low, Low]=.25 [Low High, Medium Low]=.50 [Medium, Medium High]=.75 , [High, Very High]=1 Stroke3 [Very Low, Low]=.25 [Low High, Medium Low]=.50 [Medium, Medium High]=.75 , [High, Very High]=1 Actual Data on Dataset Table 3 Dataset Pre-analysis: Input Actual Stroke Factor Variables Variable Surgery Sub-Groups & Ranges [Very Low, Low]=.25 [Low High, Medium Low]=.50 [Medium, Medium High]=.75 [High, Very High]=1 Rehabilitation [Very Low, Low]=.25 [Low High, Medium Low]=.50 [Medium, Medium High]=.75 [High, Very High]=1 Actual Data on Dataset Length Of Stay [Very Low, Low]=.25 (LOS) [Low High, Medium rehabilitation Low]=.50 [Medium, Medium High]=.75 [High, Very High]=1 Where 1=36 days of Rehabilitation Page 6 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Table 4 Dataset Pre-analysis: Target Variables to predict With dataset grouped on 4 logical groups our next step is to implement the Neural Network for the minimum possible prediction error. Implementation of the neural network: We proposed to train a NN under Supervised mode (produce output in response to sample inputs, making them particularly well-suited to modeling and controlling dynamic systems, and predicting future events). Where we have 14 Input Variables on three groups: Stroke independent, Risk Factors and Stroke Series and predict the Target variables: Surgery, Rehab and Los_rehab. As shown on Figure 2. Figure 2 Proposed Neural network to predict three stroke variables The biggest problems on Neural Network are: How many Hidden Layer and Neurons, and what Criteria? To predict the three parameters mentioned: Surgery, Treatment and Days of rehabilitation using with the actual information available. We use three different software available: NN Basic technique (with no automatic stepwise functions, limited diagnostic), another NN semi-automatic technique (allows stepwise function and good diagnostic) and one Neuro-Intelligence (use Genetic Algorithm to find the best NN configuration and an excellent diagnostic tools) our proposed solution must be minimum possible prediction error. Page 7 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa 1) NN Basic technique_ Analysis using WEKA Open Source tool for Data mining [3].Weka includes a number of different techniques that can be useful for forecast development. These include: Linear and logistic regression Perceptron Models (Neural networks) and a variety of classification algorithms are available as Standard algorithm like J48, which is version of the last free version of Classification and regression trees (CART) is a nonparametric decision tree learning technique that produces either classification or regression trees, depending on whether the dependent variable is categorical or numeric. Unfortunately, the “work horse” linear regression module in Weka is limited in usefulness: No automatic stepwise function, Poor diagnostics Compare as others software. Weka Pro and Cons are shown on Table 5. That why we didn’t choose Weka for our purpose on this research. Weka techniques for NN Different techniques as: Linear and logistic regression Perceptron models (Neural networks) Weka classifiers for NN Includes as: J-48 based on Classification and regression trees Weka limitations for NN On the Linear Regression Module: No automatic stepwise function, Poor diagnostics Table 5 Dataset Analysis using WEKA Tool to predict Stroke Variables Page 8 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa 2) NN semi-automatic technique that allows stepwise function and good diagnostic. We choose Matlab Toolbox [2]. We based our topologies of NN based on previous studies for Roy, Cheng, Chang, Moore De Luca[5]. The choose to analyze four different topologies: Two with a Single hidden layer 22 and 33 neurons respectively and two NN with Two hidden layers 44 and 33 or 44 and 22 neurons respectively. Where each of the three targets was assigned to one output neuron. This is an important characteristic of the design resulting in orthogonal outputs that decrease the likelihood of misclassification errors. They recommend a Tan-sigmoid transfer function was used for the neurons of the hidden layers, and a linear transfer function for the neurons of the output layer as shown on fig.3 Figure 3 Proposed Analysis of Four Neural networks to predict three stroke variables To define the training function of the Neural network we have the option to choose three transfer functions: Logsig (generate output between 0 and 1), Tansig (multilayer networks) usually it is recommended for recognition problems and Purelin (linear output neurons) recommend for fitting problems, as shown on table 5. Page 9 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Table 5 Options available for transfer functions on NN For training the proposed NN we have different algorithms as shown on Fig. 4, each one with Adv and disadvantages for the kind of data that we need to train. Roy, Cheng, Chang, Moore De Luca [5] recommend the use of Scaled Conjugate Gradient SGM algorithm for solving the propose NN because SGM deal with weights and biases of neurons it is a supervised learning algorithm optimization class techniques well known in numerical analysis as the Conjugate Gradient Methods, besides testing with different algorithms we found the Bayesian regularization algorithm BR results very closed to the results of the analysis of SGM. Page 10 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Figure 4 Algorithms available for train the proposed NN For evaluate the results we use two criteria’s: Performance chart (should get a decaying plot, since you are trying to minimize the error) and Regression Parameters chart verifying the parameter R to be as high as possible. We form two groups for testing one based on the using Scaled conjugate gradient algorithm SCG results on table 6 and the other on Group B using BR Bayesian Regularization table 7. Page 11 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Table 6 Evaluation NN results using SCG as training Algorithm Table 7 Evaluation NN results using BR Bayesian Regularization Analyzing the results we detect the optimal configuration was with the Bayesian Regularization training method with a Performance curve showing decaying plot and The Regression Value =.79 (21% forecasting error ), but can we find a better solution for the proposed NN? To answer this question we use another technique on the next step. Page 12 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa 3) Neuro-Intelligence that use Genetic Algorithm to find the best NN configuration and an excellent diagnostic tools our proposed solution must be minimum possible prediction error. We use Alyuda Neurointelligence software because we have tool for Analysis, pre-processing, Architecture search algorithm to recommend the best NN configuration. Running the option on search NN best architecture the recommendation was a network topology with 1 hidden layer with five neurons but for the inputs a configuration of 30 connected to the 14 stroke variables and for output 7 connected to the 3 stroke target variables as shown on fig. 5. Fig. 5 Recommend NN architecture for the proposed Stroke NN Running the training on Neuro-Intelligence we validate the results that we get using on NN semi-automatic technique using the Conjugate Gradient Descent to a value of 78 (22 % forecasting error) ,other algorithm recommend was the Batch Back Propagation and we get an improvement to validation=.81 ( 19% forecasting error) as shown on table 8. Page 13 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa NN Topology 14 Input to 30-5-7 3 Outputs Training Algorithm Batch Back Propagation Dataset Error s 14 Input to Conjugate 30-5-7 Gradient 3 Outputs Descent Table 8 Validation of NN with topology 30-5-7 with two different algorithms The NN recommended topology of 7 Inputs to 30-50-7 3 Outputs using Bach Back Propagation for training and verifying the Output with the Data Target the results on a chart are shown on Fig. 6 Fig. 6 Verifying Actual Target vs Output from the NN 30-50-7 Page 14 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Experimental results: Post-Validation of the neural network Post-validate the proposed NN of 30-5-7 on two ways: 1) Post-Validation using new input data with a logical values and expected logical values as a target. For example we expect that a healthy subject from 45-55 years old, active, with no Stroke Risk Factors will forecast a values of zero for: Surgery, Rehab and Los_Rehab, and a man from 76-86 year old, active with all the Risk Factors and three strokes will get a very medium value for Surgery, Medium High for Rehab and Los_Rehab. These results are Shown on Table 9 2) Post-Validation using Random data generated with a uniform distribution generate for the program itself. We verify the results and they are shown on table 10. Table 9 Post-Validation new input data with a logical Target values Page 15 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa Table 10 Post-Validation new Random input values and forecast expected The post-validation results where according with the values that we expected. The show that the proposed NN of 30-5-7 it is working and forecasting the expected values for: Surgery, rehab and Los_Rehab. Conclusions: Specific Conclusions: Based on the 14 Stroke input variables and the 3 output target Stroke values, this paper suggest that the forecasting of : Surgery, Rehab and Los_Rehab it is possible using Neural network, a bigger Dataset is recommended to achieve an optimized NN to have a minimum of 10% of forecasting error. General conclusions: ANN’s are a powerful technique utilized across scientific disciplines. Theoretically well suited to non-linear processes like this application, but NN it is not transparent to users, Hard to integrate into forecast thinking with technically difficult to understand, raises risk of misuse. Further works: Create a larger dataset with own lab reading values to have a better control of the information and data to have realistic results and apply a follow-up the subjects during its recovery with an the goal of feedback and take appropriate forms to improve this recovery, and connect these results with a complete Recovery systems for the Follow-up of the subject and achieve the Database Acknowledgment (DBK) Based on: Surgery, Treatment and Days of rehabilitation to take a decision based on a System Design for Follow-Up of subjects with Mild Stroke to feedback Medical Doctors if the subjects are responding to treatment for a faster recovery. Page 16 Artificial intelligence analysis using Neural Network to predict three stroke parameters: Surgery needed, Treatment, and Length of Stay for Rehabilitation by Jorge Garza-Ulloa References: [1] Stroke_full_database from Many Eyes An experiment brought to you by IBM Research and the IBM Cognos software group on the link : http://www-958.ibm.com/software/data/cognos/manyeyes/datasets/stroke-fulldataset/versions/1 [2] Neural Network Toolbox™ Mathworks (2009) [3] WEKA Open Sources tools for Data Mining; http://www.cs.waikato.ac.nz/ml/weka/ [4] NEURAL NETWORKS: Heikki N. Koivo (2008) [5] Neural Network Biomedical Engineering IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 17, NO. 6, DECEMBER 2009 by Serge H. Roy, M. Samuel Cheng, Shey-Sheen Chang, John Moore, Gianluca De Luca,S. Hamid Nawab, Carlo J. De Luca. [6] J. R. Lieberman, F. Dorey, and P. Shekelle, “Difference between patients’ and physicians’ evaluations of outcome after total hip arthroplasty,” J. Bone Joint Surg., vol. 78, pp. 835–838, 1996. [4] D. B. Reuben, “What’s wrong with ADLs?,” J. Am. Geriatr. Soc., vol.43, pp. 936–937, 1995. [7] K. Kiani, C. J. Snijders, and E. S. Gelsema, “Computerized analysis of daily life motor activity for ambulatory monitoring,” Technol. Health Care, vol. 5, pp. 307– 318, 1997. [8] K. Kiani, C. J. Snijders, and E. S. Gelsema, “Recognition of daily life motor activity classes using an artificial neural network,” Arch. Phys. Med. Rehabil., vol. 79, pp. 147–154, 1998. [9] ] D. M. Sherrill, P. Bonato, and C. J. De Luca, “A neural network approach to monitor functional motor activities,” presented at the 2nd [10] M. S. Cheng, “Monitoring functional motor activities in patients with stroke,” Ph.D. dissertation, Boston, MA, 2005. [11] A. R. Fugl-Meyer, L. Jaasko, I. Leyman, S. Olsson, and S. Steglind, “The poststroke hemiplegic patient. I. A method for evaluation of physical performance,” Scand J. Rehabil. Med., vol. 7, pp. 13–31, 1975. [12] R. A. Keith, C. V. Granger, B. B. Hamilton, and F. S. Sherwin, “The functional independence measure: A new tool for rehabilitation,” Adv. Clin. Rehabil., vol. 1, pp. 6–18, 1987. [13] J. Mao and A. K. Jain, “Artificial neural networks for feature extraction and multivariate data projection,” IEEE Trans. Neural Networks, vol.6 Issue:2 1995 Page 17