We propose to develop a framework for an intelligent reasoner with capabilities that support complex decision making processes in medical diagnosis. Identifying the causes, reasoning the effects to explore information geometry and learning the associated factors, from medical forum information extracted, are the core aspects of this work. As part of the proposed framework, we present an approach that identifies semantically similar causes and effects for any specific disease from medical diagnosis literature using implicit semantic interconnections among the medical terms. First we crawled MedHelp1 forum data and considered two types of information: forums data and posts data. Each forum link points to a specific disease and consists of several topics pertaining to that disease. Each topic consists of multiple posts that carry either users' queries/difficulties or doctor's feedback pertaining to the issue(s) of the users. We use graph based exploration on the terms (diseases) and their relations (in terms of causes/effects) and explore the information geometry pertaining to similar diseases. We performed a systematic evaluation to identify the relevance of the contextual information retrieved for a specific disease and similar factors across different diseases. The proposed approach looks promising in capturing similar causes and/or effects that pertain to multiple diseases. This would enable medical practitioners to have a multi-faceted view of a specific disease/condition.
Understanding the causes and effects pertaining to a specific disease is key to better prediction in medical diagnosis and improved patient management. Diseases/Conditions may have similar or semantically related causes and effects. Furthermore, gaining insight on how diseases are diagnosed and managed would enable 1http://www.medhelp.org/forums/list medical practitioners to make more informed decisions on disease management. Illustrating the role of machine learning as an enabler of this for more informed decision support is a key aspect of this paper. In this paper, we present an approach to identify causes and effects that exist across different diseases using a graph clustering based knowledge discovery approach.
Understanding causation and correlation between individual factors is key to decision making in multiple domains including medicine and business. Traditionally, such association between elements has been identified through human interpretation of text. However, this is a very time consuming, manual, labour intense process and is limited by human capacity. The ability to mine textual content, identify association and the nature of that association between elements using machine learning techniques provides significant opportunities. Specifically in the medical domain, where a significant amount of content is textual in nature (e.g. medical notes), having the ability to identify causation and correlation between elements in large medical datasets provides significant opportunity for advancing medical research and enabling better decision making pertaining to condition diagnosis and patient management.
In this paper, we attempt to identify and create term clusters using graph clustering approach and then perform topic classification. This approach improves the document classification task by putting the terms, that are semantically related, in the same cluster. Users could search for a specific disease and explore information pertaining to its causes and effects by means of semantically related texts.
To incorporate natural language understanding,
commonsense and domain specific knowledge could be used to
improve the text representation by including more generated
informative features to perform deep understanding of the
document text than the mere Bag-of-Words approach [
In this section, we characterize the medical forum data in a formal way. Each post pertaining to a specific topic is informally written and we focus on terms and their cooccurrences. We have n textual descriptions, viz-a-viz, posts: P = {p1, p2, · · · , pn} and each post can be formally represented as a sequence of terms, illustrating the scenario of the underlying disease, as follows: pi = {t1, t2, · · · , tm} where 1 ≤ i ≤ m and m varies differently for different post. We convert the entire text data of all posts into a graph, say G = (V, E) where V represents the set of nodes (each term in P is considered as a node) and the co-occurrence of any pair of nodes across the posts is considered to be the edge representing the strength of association between the pair of nodes. 2.2
Clustering deals with identifying a pattern/structure in the bunch of unlabeled data. In general, clustering organizes data into groups whose members are related in some way and two or more data can be grouped into the same cluster if they are, in some way, falling close to each others’ context. Clustering has many useful applications like finding a group of people with similar behavior, processing orders, grouping plants and animals, grouping web blog data to access similar patterns.
While exploring a variety of possibilities to identify the context of causes and effects of diseases from text fragments, feature space grows and it is hardly possible to limit the expansion of the new feature space containing the local contexts extracted from the informal writing of medical data. This could possibly be solved by using dimensionality reduction techniques to limit the size of the document to be classified. This attempt first makes word cluster vectors using unsupervised feature generation by identifying the related contexts. Then using the identified contexts, supervised learning is performed for categorizing the given text collection consisting of user posts.
First, we use the entire collection of medical diagnosis related posts and filter out the list of the distinguishable unique terms. Using these terms, we first build the weighted graph in which nodes represent terms and edges represent the weight - the number of documents in which the given pair of terms co-occurs across the collection of posts. For 2word and term are used interchangibly each unique term, the list of documents in which it occurs is retrieved. Using this data, we build the weighted graph in which the edge between two terms would represent their semantic association implicitly. This process is repeated for all features and a weighted graph for the overall data is constructed. Thus the problem is modeled into a graph clustering problem. This results in a graph G = (V, E, A) where |V | = n represents the number of unique terms; |E| represents the number of edges and the adjacency matrix; and A is |V | × |V | whose nonzero entries correspond to the edge weight between a pair of terms (adjacency list is assumed in case of sparse matrix - in this case, number of rows in the graph represents the total number of terms in the graph).
We use the kernel-based multilevel clustering algorithm
proposed by Dhillon et al. [
In this work, we generate term clusters from extracted word graphs, using co-occurrence information of terms. The task is to partition the graph into clusters so that terms could be grouped into a few subsets and dimensionality of the new term space is reduced. Now based on the generated term clusters, we perform classification to identify similar causes and effects across various diseases. 2.3
The proposed approach works as follows: From the posts of each topic, textual descriptions are extracted. Unique terms (after removing stop words) are considered as nodes in the graph and the number of times a pair of terms cooccurs in the entire corpus is considered as the weight of the edge connecting the pair of terms. At first, we build word cluster vectors using graph clustering algorithm. Algorithm 1 Building Word Clusters Input: A set of n textual descriptions (posts)
P = {p1, p2, · · · , pn} A set of predefined category labels C = {c1, c2, · · · , cl} Build Word Cluster Vectors: 1: Extract text from posts and build the unique word list 2: for each unique term ti in the word list do 3: Identify the existence of edges from ti to all other terms with nonzero positive weight. 4: Store the co-occurring term with its corresponding edge weight in the adjacency list 5: end for 6: Use kernel-based multilevel graph clustering algorithm on the adjacency list and perform clustering to generate cluster IDs 7: For every cluster ID, construct word clusters 8: Store these word cluster vectors
Secondly, we use these word cluster vectors to re-represent text documents that discuss the causes and effects of a specific disease. Then we perform classification of the re-represented text documents. The proposed algorithm inherently applies clustering semantically related terms and performs classification of them. Based on the word clusters found in the first step, we perform the classification on the clustered space of label pertaining to the terms in the original documents.
Algorithm 2 Classification using Word Clusters 1: Preprocess the text documents by removing numbers, punctuations and stop-words (using SMART3 word list) 2: for each processed text (post) data pi in P do 3: for each unique term in pi do 4: Identify its cluster id 5: Map the given feature in terms of its cluster ID 6: Augment text fragments with cluster ID mappings 7: end for 8: end for 9: Build classifier on these mapped/expanded text data (containing only cluster IDs) and use it to predict the class of the text having similar causes and effects 10: Compute the classification accuracy Output: The category label(s) for texts (post) that have similar causes and effects across diseases.
We crawled a subset of MedHelp4 forum data from the world wide web. The MedHelp forum is organized as a set of topics, each representing a specific disease and under each topic, there are several subtopics. Each subtopic consists of several posts from both users (may be patients or their dependents) and doctors as well.
For our experiments, we have selected 15 categories covering the most widely discussed topics in the MedHelp forum. The details of this experimental corpus is given in Table. 1
We have used 3 different types of classifiers, namely Naive Bayes, k−Nearest Neighbours (k−NN), and Support Vector Machines (SVM), to test the effect of the proposed approach. We have used rainbow5 to build the classification models and during classification, we have used 60% of the data for training and 40% for testing. 3.1.1
Evaluation Methodology
We have used Precision, Recall, F-Measure and the classification Accuracy to evaluate the quality of the identified diseases having similar causes and effects. We use the two-way confusion matrix given in Table. 2 to derive the evaluation measures:
Precision is defined as follows:
Class Asthma Breast-Cancer COPD Cosmetic Dental Embarazo Genetic-Disorder Hepatitis Kidney Liver-Transplant Oral Pathology Respiratory Thyroid-Cancer Varicose-Veins #Diseases (Posts) 35 (43) 35 (37) 41 (47) 57 (64) 97 (100) 41 (52) 59 (68) 110 (147) 69 (89) 64 (62) 35 (47) 31 (36) 48 (54) 58 (63) 36 (52) 80.00 cya ru60.00 cc A 40.00 20.00 0.00
We have observed the classification accuracy of k−NN classifier for various values of k (=10, 20, 30, 50) and found that for k =30, the system performs very well. In this classification task, we have observed that the maximum number of instances from the class “Liver-Transplant” is misclassified under “Hetatitis” class as the causes of these common diseases coincide by and large. At the same time, none of the instances is classified correctly for the class “Pathology” as these causes and effects are very similarly described as that of “Hepatitis” and “Thyroid-cancer”. While using the SVM classifier, we noticed that some instances are misclassified across the classes: “Breast-cancer” and “Cosmetic”. In this case, user raised cross reference related queries with post surgical treatment of the Breastcancer disease. Similar misclassification is found across the classes: “Chronic Obstructive Pulmonary Disease” (COPD) and “Resporatory”. Subsequently, we would like to apply this approach for effective retrieval of causes and effects pertaining to a specific disease. Also we will draw the information geometry of prominent diseases that share the common causes/effects in our subsequent experiments.
We proposed a method to enable greater understanding of various conditions, their symptoms, treatment and management by identifying similar scenarios, reasoning the effects to explore information geometry and learning the associated contextual factors, from medical forum information extracted from health services data. This approach identifies semantically similar causes and effects for any specific disease/condition, using implicit semantic interconnections among the medical terms. We use graph based exploration on the terms and their relations (causes/effects) across the collection of posts and explore the information geometry pertaining to the similar diseases. We evaluated the relevance of the contextual information retrieved for a specific disease and/or similar factors across different diseases. The proposed approach looks promising in capturing similar scenarios pertaining to multiple diseases. This would enable medical practitioners to have a multi-faceted view about any specific disease, towards better decision making. 70.86 62.54 54.13 NaiveBayes krNN (k=30) Acknowledgments: This research is co-funded by the Irish Government (Enterprise Ireland) and European Union (European Regional Development Fund). Dr. Philip O’Reilly is the Principal Investigator responsible for this research and can be contacted at Philip.Oreilly@ucc.ie.