مطالعه درباره هوش مصنوعی و رگرسیون لجستیک برای کمک در تشخیص دیفرانسیلی سرطان لوزالمعده
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|24782||2009||10 صفحه PDF||سفارش دهید||6700 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Expert Systems with Applications, Volume 36, Issue 7, September 2009, Pages 10663–10672
Recent medical researches indicate that pancreatic cancer is the eighth leading cause of death of the populations in Taiwan. Each year approximately 800 victims die of this cancer, and the number is increasing year by year. Since most early symptoms of pancreatic cancer is non-specific, doctors’ diagnostic decisions might differ based on individual experience, knowledge of the disease, and influenced by their mental conditions at examinations. Certain diagnostic errors are inevitable to occur and mislead the following treatment plans and thus result in insignificant and inefficient follow-up tests. This phenomenon not only caused wastes of the medical resources but also severely delay the golden timing of “early detection and early treatment” for patients. This study used artificial neural network and genetic algorithm of artificial intelligence (AI) as well as logistic regression of statistics to construct three types of screening models for pancreatic cancer and acute pancreatitis. Additionally, it adopted the ROC curves to compare and analyze discriminations of the above-mentioned three screening models. It used 234 case patient data as its training samples and 117 cases as test data. The results of pairwise comparisons and analysis indicate that AUC values of the tree models have no significant differences regardless. Except the fact that GALR model is obviously better than SLR model, both the pairwise comparisons between SLR and BPN or BPN and GALR have no significant difference. On the contrary, however, if under the condition of obtaining the optimal threshold of the three models, GALR model has the best performance with 96.7% in sensitivity and 82.5% in specificity, which are both better than SLR model with 96.7% in sensitivity and 73.7% in specificity and BPN model with 88.3% in sensitivity and 84.2% in specificity. Finally, artificial intelligent approaches will have more optimal predictions in the future with larger and more comprehensive data base as well as more accurate computing methods.
The pancreas is an important body organ that has both exocrine and endocrine functions. Since the pancreas is located in the posterior of the abdomen, signs and symptoms of pancreatic cancer strike at a slower pace and often go unnoticed before it spreads. Therefore, comparing with other types of cancers, regardless of European or oriental countries, prognosis of pancreatic cancer is generally regarded as poor, the causes are the hardest to found, and it is the type of abdominal cancer that has the worst survival rate. Clinically when physicians encounter patients with symptoms similar to both acute pancreatitis and pancreatic cancer, diagnosis is made based on individual experiences. First, a doctor may initially suspect the patient with a certain disease. Following the hypothesis, more tests like the endoscopic procedures, imaging tests or blood tumor markers will be prescribed by the doctor to obtain further information of the disease before making final assessment and confirming diagnosis. However, diagnosis results still might differ depending on the mental and emotional conditions of the doctor. The possibility of false diagnosis is still high. From past literatures in the related filed of both international researches and local studies in Taiwan, many cases of pancreatic cancer were initially diagnosed as other abdominal diseases due to its non-specific symptoms at early stage and thus making treatment plans ineffective (NTUCM Clinicopathological Conference, 2004). To date, there is no effective screening method focusing on early stage of pancreatic cancer nor is there a specific diagnostic tool that has sufficient sensitivity (Chen, 1997). Lacking a proper diagnostic tool at the early stage, the following-up diagnostic screening and treatments might not reach satisfactory recovery results, either. Therefore, if a series of intelligent computer supplemental system can be established to assist in diagnosis and providing references and/or suggestions for a doctor to conduct follow-up tests tailored to a specific cause and to make correct diagnostic decisions, human judgmental errors will decrease, patients will receive better treatments, and most importantly, quality care is enhanced. As a result, this study attempts to use AI to establish an auxiliary diagnostic screening model for pancreatic cancer. Mainly, this model is to help clinical doctors, when facing patients with abdominal diseases, take advantage of the high-performance computing technology of computers and use the huge volume of information stored in the database to build a suitable diagnostic model based on symptom information, physical examination results, and lab tests of patients. Then, as the next step, the doctors can prescribe follow-up endoscopic procedures, imaging tests, or tumor marks as a referential indicator to reduce man-made judgmental mistakes and eliminate wastes of medical resources.
نتیجه گیری انگلیسی
In this study compared the SLR model in the traditional statistics method, the BPN model in the artificial intelligence approach, and the GALR model integrating both AI and the statistics. The purpose is hoping to objectively provide an effective supplementary tool and provide important references and index systems for clinical physicians to prescribe follow-up endoscopic, imaging, or blood tumor marker diagnosis, etc as the next step to reduce human errors and prevent medical wastes. However, since the proposed tool is limited to provide prognostic reference, final judgments could only be reached in combination with comprehensive examinations of the specialists. The screening diagnostic model for pancreatic cancer and acute pancreatitis obtained from the research results show that age of the patient (X2), whether or not the patient shows symptoms of jaundice (X10), and patient amylase level (X40) are important indicators of the screening diagnostic model. After all, most of the previous studies on the screening diagnostic model of pancreatic cancer and acute pancreatitis in the past are only limited in imaging diagnosis, blood tumor marker, gene mutations, etc. Few are focused on population statistics data, clinical symptoms and physical examinations, or medical histories as research variables. Regarding population statistics data and clinical symptoms, the screening diagnostic model of this study shows that patient age (X2) has significant predictions and presents positive correlation with the onset of pancreatic cancer. In addition, jaundice (X10) is also positively associated with the occurrence of pancreatic cancer. Besides population statistics data and clinical symptoms factors, it is simultaneously found that amylase level among the lab test variables shows negative correlations with pancreatic cancer. Such a finding is both meaningful in public hygiene issues and the preventative medicine field. The screening diagnostic models in this study, through the process of experiments, are found to have good screening ability using artificial intelligence to build the supplementary screening diagnostic model for pancreatic cancer. Among all, the AUC value of GALR model is 0.921; that of the BPN model is 0.895. Furthermore, since GALR model is significantly better than the SLR model based on the conventional statistical method, with AUC value of the SLR being 0.882. Although comparing results of the AUC values cannot differentiate significantly in statistics between BPN, GALR, or SLR, if obtaining an optimal criterion value from the three models, the comparisons of the obtained sensitivity and specificity then show that GALR has the best screening ability following by SLR model and the BPN model last on the list.