For accurate clinical decisions, doctors may sometimes have to consult external information for reference. The published biomedical articles, which are expert-written materials that cover nearly all topics in the medical area, are the most common sources of reference [2, 3, 25]. Many biomedical literature search engines, such as PubMed1, have been developed to facilitate this data access task. However, most platforms only support keywords retrieval, thus impose a high requirement for the doctors to construct their queries accurately.
However, it is often hard for doctors to generate accurate queries because their information needs are often complicated and exploratory. For example, Ely et al.  identified that doctors’ clinical questions could deal with the following three aspects simultaneously:
Q1: What is the patient’s diagnosis?
Q2: What tests should the patient receive?
Q3: How should the patient be treated?
Because doctors often only have limited information about the patient’s current condition, such as symptoms or disease history, they usually find it hard to generate an accurate query. Therefore, better search support technologies are critically needed in such scenarios.
In order to provide more effective biomedical text retrieval technologies to doctors, Text Retrieval Conference (TREC) hosted Clinical Decision Support (CDS) tract2 between 2014 and 2016. The participants of this track have worked on retrieving relevant biomedical articles to answer the above-mentioned three questions with 90 sample electronic health records (EHRs) [2, 3].
The outcomes of TREC CDS provide two important insights. The first one is that terms extracted from EHRs alone are too ambiguous to reflect the true information needs. For example, Balaneshin-kordan et al.  found through their 2014 CDS participation that non-relevant documents talking about wrong diseases were returned when the queries contained evidences only from the EHRs. This is because the provided EHRs data only contains some incomplete disorder-related information such as disease history, symptoms, and testing results (Figure 1). However, the same symptoms and disease history information might partially be shared by different diseases. An example is that patients of either hypothyroidism or hyperthyroidism might have symptoms of dyspnea, hair loss, and fatigue. Thus, extra information is needed to enrich the queries.
The second insight is that correctly identified diagnosis, which clearly states the possible disease, would significantly improve the retrieval performance. For example, 2015 TREC CDS had task A and task B, both of which have 30 patients’ EHRs, but 20 of the 30 EHRs in task B were also provided with diagnosis information. The evaluation results showed that both the median and mean performances of task B were increased by 8% when compared with that of task A, reflecting the benefits of having the diagnosis information in helping biomedical literature retrieval .
These above-mentioned findings motivated us to conduct diagnosis predication before generating queries for helping doctors in their biomedical text retrieval. Particularly, we are interested in obtaining medical diagnosis information from publicly available knowledge bases because large quantity of EHRs is usually not available openly.
In this paper, we explore two types of large-scale open knowledge bases, which cover information about wide range of diseases and related information, such that they are suitable to be mined for possible diagnosis-related information. The first one is Wikipedia, which represents the type of open corpus of free text. Wikipedia presents information at word level, with rich context information about each disease and its related symptoms, tests, and treatments. The second one is Semantic MEDLINE Database (SemMedDB), which consists of medical concepts extracted from PubMed3 literatures. Different from Wikipedia, SemMedDB organizes its information around medical concepts and their relationships. This concept-oriented expression can be useful because doctors, medical literature, and Wikipedia articles might express the same concepts but with different words or phrases. However, SemMedDB has its own limitations. Kilicoglu et al.  reported that NLM’s SemRep, the tool used to build SemMedDB, only achieves 75% extraction accuracy. This means that SemMedDB itself contains many noisy extraction results. Consequently, our method uses the combination of Wikipedia and SemMedDB to extract diagnosis-related information to support doctors’ biomedical literature retrieval.
Once the diagnosis is predicted, we view the rest of the method as a query expansion problem. The expanded query contains the original parts that are generated from the disorder-related medical concepts recognized in the EHRs, and the expanded part is the predicted diagnosis. Therefore, our model consists of five modules: medical concept extraction with MetaMap, Wikipedia-based diagnosis predictor, SemMedDB-based diagnosis predictor, prediction fusion, and query expansion with diagnosis.
There are quite a few works studying the large-scale diagnosis prediction, so we compare our work with those in this paper. Since we have no correct diagnosis, we extrinsically validate our method on CDS retrieval performance.
The remainder of this paper is presented as follows. Section 2 shows the related works. Section 3 describes our proposed methods. Sections 4 and 5 give the experiments and discussions. Finally, Section 6 presents the conclusion and future work.
2 Related Works
The related work can be divided into two parts. The first part examines the existing studies on achieving automatic diagnosis prediction, and the second part talks about query expansion in medical retrieval.
2.1 Automatic Diagnosis Prediction
Automatic diagnosis prediction is a very popular research topic that attracts many researchers. Esfandiari et al.  gave a comprehensive review of the studies in this area. According to their summarization, most studies regarded diagnosis prediction as a classification task. For example, Yeh et al.  used the history of patients’ diseases, blood test results, and physical examination results as the features and used trained classifiers to predict the probability of getting a cerebrovascular disease. Besides, other studies tried to predict the diagnosis with regression methods, clustering methods, association rules, or hybrid systems . For example, Briones et al.  tried to predict the risk of Alzheimer disease through a regression model. However, these works usually targeted on sensitively and accurately predicting a small range of diseases. For instance, Yeh et al.  concentrated only on the cerebrovascular disease prediction and obtained 98.01% accuracy and 94.68% sensitivity.
There have been a few works exploring the large-scale auto-diagnosis. Isola et al.  proposed a neural network and a KNN-based system to predict diseases. But their work focused on the implementation of such a system and did not provide performance data on the prediction. Gomathi and Nithya  constructed a well-structured database to compute the probability of a disease based on a patient’s symptoms, but did not provide prediction performance. Liu et al.  proposed a quite interesting android platform, allowing users to interactively communicate with the autodiagnosis application, and used user’s explicit feedback in computing the probability of the diagnosis. Nie et al.  proposed a deep learning-based framework to predict large-scale diseases for users on Question–Answer (QA) platforms. They trained the network on collected QA data and carefully refined the network on different disease prediction with precisely collected disease-related QA data. They validated their system on predicting 20 diseases and found that their method outperformed KNN, SVM, Decision Tree and Naïve Bayes. Although their system worked on large-scale disease prediction, their system highly relied on good quality training data. Koopman et al.  found that, when the training data are imbalanced, their system could encounter the classification/prediction failure for the rare but important diseases that might have little data for training.
In comparison to the related work on this part, our methods work on predicting wide range of diseases. This is because doctors’ medical literature search can be on all sorts of diseases so that our automatic diagnosis prediction system cannot afford to work only on a small number of diseases. Furthermore, we concentrate on proposing a diagnosis prediction system with no need on large quantity of training data. Finally, in contrast to past studies that only accepted well-formatted features, our system can deal with noisy medical free text. Overall, our goal is to make the prediction system quite robust and need little effort when changing scenarios.
2.2 Query Expansion in Biomedical Text Retrieval
In the generic information retrieval area, query expansion is a common module to enhance the original search, but it may lead to query drift if not designed carefully [4, 5]. In the CDS task, nearly all previous studies utilized query expansion to enhance the original query [6-10].
Most of them extracted expanded terms from Pseudo Relevance Feedback (PRF) documents within the collection [6, 8]. For example, Choi et al.  presented the best result in CDS task of TREC 2014. They utilized the most frequent Medical Subject Headings (MeSH) terms that label the PRF documents to expand the original query, and then used the classifiers to re-rank the returned document list. Both steps showed significant improvement. Balaneshinkordan et al.  expanded their query with terms selected from both PRF files and Google search results.
There were also other works using important medical concepts extracted from external resources to expand the query [7, 8, 10]. Oh et al.  proposed to enhance external expansion model (EEM) with cluster-based document model (CBEEM) to expand the query more accurately. Their PRF model consisted of top-ranked documents both in the target dataset and in the external collection, Wikipedia. Their results outperformed the best runs in CDS task of TREC 2014. Song et al.  extracted the most frequent MeSH terms appearing in the Google search result returned with the original query.
Past works explored different retrieval models in biomedical text retrieval task. Both Balaneshin-kordan et al.  and Xie et al.  used the Markov Random Field (MRF) model and got very high retrieval performance. Song et al.  proposed to retrieve relevant biomedical articles by combining three retrieval models, including BM25, PL2, and BB2, and their results performed the best in 2015 CDS task B.
In the query expansion procedure of all these previous studies, it is showed that the quality of the expanded terms is important. Since the diagnosis can better reflect users’ true information needs, we believe that adding the diagnosis predicted with Wikipedia and SemMedDB to the original query can help to clarify the original query and does not introduce too much noise.
3 Our Methods
As stated, in this paper, we propose novel methods to enhance the clinical decision system by automatically predicting patients’ disease with the publicly available online knowledge bases. To assist doctors in their clinical decisions on a wide range of diseases, the diagnosis prediction methods we build here can make predictions for large number and wide scale of diseases rather than concentrating on a small number of narrowly defined diseases. Therefore, our diagnosis prediction algorithms need to draw evidence and help from large-scale, publicly accessible data and knowledge bases, and our methods concentrate on two knowledge bases, Wikipedia, the free text collection that provides rich contextual information but at the word level, and SemMedDB, the conceptual level knowledge bases consisting of the medical concepts and their relationships extracted from PubMed. Figure 2 shows our methods, each of which integrates some or all of the following functional modules:
Medical concepts extractor, which automatically identifies the medical concepts in the given EHRs for a search task.
Wikipedia-based diagnosis predictor, which utilizes Wikipedia knowledge to predict the most probable disease diagnosis based on the EHRs of a search task.
SemMedDB-based diagnosis predictor, which utilizes SemMedDB knowledge to predict the most probable disease diagnosis based on the EHRs of a search task.
Fusion-based diagnosis predictor, which combines the ranking list of above two predictors to get most probable disease diagnosis.
Query expansion with diagnosis, which expands the original query with the predicted diagnosis.
3.1 Medical Concept Extractor (MCE)
Following previous literature [6-10], the medical concept extractor (MCE) module relies on Unified Medical Language System (UMLS) vocabulary to identify medical concepts in EHRs. UMLS is a knowledge base published by US National Library of Medicine (NLM), which contains a full list of medical concepts . Each UMLS concept has a Semantic Type (ST) attribute. For example, concept “Dyspnea” has its ST as “sosy”, and the ST of “woman” is “popg”. In this work, medical concepts with following STs are kept: acab, anab, bact, bdsu, blor, bpoc, bpoc, celc, cgab, comd, diap, dsyn, emod, euka, fndg, fngs, food, ftcn, hlca, inpo, lbpr, lbtr, menp, mobd, neop, ortf, patf, phsu, sosy, tisu, tmco, topp, virs. These STs are selected because such medical concepts provide important patient disease-related information and frequently appear in EHRs.
The implementation of MCE relies on MetaMap to extract the concepts from EHRs. MetaMap is a popular medical text mining tool developed by NLM . Met-aMap can detect the negation expression. For example, “trauma” will be ignored in “she has no history of trauma”. Our extractor module ignores negation expressions, and assigns a unique medical concept identifier - Concept Unique Identifier (CUI) in the UMLS Metathesaurus - to each extracted medical concept.
3.2 Wikipedia-based Diagnosis Predictor (Wiki-DP)
Wikipedia-based diagnosis predictor (Wiki-DP) module relies on information extracted from Wikipedia for diagnosis prediction. As an open and rich knowledge base, Wikipedia covers wide range of diseases and their related information, which, in most cases, is sufficient to act as an external resource for biomedical retrieval. Typically, a wiki page titled with a disease name would contain information about the causes, symptoms, pathophysiology, and diagnosis of the disease. Through such information, we can calculate the co-occurrence between a disease and certain symptoms, which can then be used to create models for ranking possible diseases based on the extracted symptoms.
3.2.1 Initial Query Composition
After the medical concepts are extracted, we can construct a query, which will be used in disease predictor. The query is constructed with MRF model . Works in [8, 9] also use MRF model in medical retrieval. In MRF, query consists of cliques, for example, EHR query consists of extracted medical concepts. If a clique contains several terms, term dependence information can help retrieve the relevant documents. Such term dependence in one clique can be described as terms appearing in the document in an ordered/unordered sequence within a window size. Given a query Q, document D can be ranked as:(1)
where T is the unigram clique set in the query, which has no term dependency; O is the cliques of ordered terms having sequential dependency; U is cliques of unordered terms having sequential dependency; λT, λO, λU are weighting parameters for unigram, ordered cliques, and unordered cliques, respectively, and they add upto 1, and f(D|c) is the probability of the document appearing given the clique. For example, for an unordered clique “chest pain”, fT(D|c) can be described as “#uw(chest pain)” in Indri language .
In this work, according to the experiment results of training data, we consider only the unordered term dependency and independent unigrams. Each extracted medical concept is a clique. In Indri query language, such query can be written as:(2)
where λT + λU=1. For example, after processing the summary text in Figure 1, we have a set of terms: breathing difficulty, dysphagia, fever, drooling, stridor, and voice change. Thus, the query is:
#weight(λT #combine(breathing difficulty dysphagia fever drooling stridor voice change) λU #combine(#uw(breathing difficulty) dysphagia fever drooling stridor #uw(voice change)))
3.2.2 Ranking the candidate diseases
We assume that the disease (predicted diagnosis) cooccurs frequently with its symp-toms in the Wikipedia articles. Therefore, using the above obtained query, we can retrieve the most relevant Wikipedia articles, and expect that the most relevant wiki pages contain the diagnosis. We downloaded Wikipedia and index it with Indri (See more details in next section). Although there could be many Wikipedia articles talking about entities in other domains, such as foods, traveling, and policy, these noisy entities do not bother our predictions. This is probably because few terms are shared be-tween the query and the unrelated entities. Indri evaluates document relevance by language modeling with Dirichlet smoothing, in which we have a smoothing parame-ter μ defining the degree to overcome data sparseness and ‘zero-probability’ problem .
After wiki pages are returned, MetaMap is used again to identify if the article title is a disease name. If yes, this title is selected as diagnosis. If not, that article is ignored and the next article is considered. In total, top 10 wiki articles participate in providing possible diagnosis.
3.3 SemMedDB-based Diagnosis Predictor (SMDB-DP)
Wiki-DP draws disease information based on a search to the free-text content of Wikipedia, but SemMedDB-based Diagnosis Predictor (SMDB-DP) utilizes SemMedDB, which is a repository of concepts and their relationships extracted from PubMed using a tool called SemRep4. SemMedDB is a network of medical concepts, with concepts represented as nodes and the co-occurrence of concepts in the medical literature represented as edges. Similar to Wiki-DP, SMDB-DP also assumes that the true diagnosis should be the disease that co-occur frequently with the extracted symptoms (or signs or history diseases). Therefore, the disease concepts in SemMedDB can be ranked based on their co-occurrence with the extracted symptoms.
However, for SMDB-DP to achieve good performance in diagnosis prediction, two important issues should be resolved. The first one is about partial matching of extracted symptoms and those mentioned in SemMedDB. Because all extracted medical concepts might not cooccur with a disease in one medical article, it is common that only parts of the extracted symptoms are mentioned together with a disease in one document. Therefore, partial matching needs to be handled.
The second issue is that, although it would make the model much simpler, the symptoms associated with a disease cannot be regarded independent to each other. The risk of viewing these symptoms to be independent is that it can cause severe topic drift (disease drift) problem. For example, popular disease can appear in thousands of papers, such as fever appears in 126,396 articles, while rare diseases are very infrequent, for example, voice change appears only in 51 articles. If fever and voice change are independent of each other, the predicted candidate disease will be dominated by fever-related popular diseases.
Hence, suppose n medical concepts are extracted from the medical record, and we assume that they are dependent on each other, SMDB-DP would consider a disease to be true only when at least concepts co-occur. And the probayility of a disease being the true diagnosis is calculated as:(3)
In this way, we can get a disease ranking list for each query.
3.4 Prediction Fusion
We assume that a disease is highly probable to be correct if it is predicted as true diagnosis by both SemMedDB and Wikipedia.From the results of the experiments in the next section, we found that Wikipedia has a better and more robust prediction across three datasets, and hence, we use the following rules to combine the prediction outputs from Wiki-DP and SMDB-DP:
Only top 10 diseases are considered in both ranking lists.
If the two lists share the same diseases, the shared diseases are kept and ranked with Wikipedia ranking score.
If the two lists do not share the same diseases, select the top disease in the Wikipedia ranking list.
3.5 Query Expansion with Diagnosis
After obtaining the predicted diagnosis, we can use it to expand the original query. In the format of Indri query language, this combination is shown as follows:(4)
where weighting parameters α ranges from 0 to 1. Original query is the query used in Wiki-DP module to retrieve Wikipedia articles, which contains patient symptom information but without diagnosis.
We conducted a set of evaluations to validate the effectiveness of our proposed method.
4.1 Dataset and Metrics
Our study included five datasets (Table 1). Two datasets were used for building diagnosis prediction algorithms. The English Wikipedia collection (enwiki)5 was used for Wiki-DP. It was downloaded on March 5, 2016, and contains 5.79 million articles. Only the title and the content of each article were kept. Tags, references, external links and see also parts were all removed. The Wikipedia collection was first performed to stop word removal and stemming using Porter stemmer, and it was then indexed by Indri.
The SemMedDB contains medical concepts extracted from 26.7 million PubMed citations. We downloaded the PREDICTION table, which was published on December 31, 2016. This table contains the CUI pairs appearing in each PubMed article.
The three data collections used for evaluating our method were from the CDS track of TREC 2014, 2015, and 2016. The data collections for the CDS track in TREC 2014 and 2015 contain the same set of 744,138 articles from PubMed Central. The data collection for CDS in TREC 2016 has 1.25 million articles. Each article in these three collections contains only title, abstract, and article content. The CDS track in each year provided 30 patients’ EHRs as the search topics, each of which, as shown in Figure 1, has three elements: description, summary, and either diagnosis in 2014/2015 or notes in 2016. To compare with past studies, we extracted the query from the summary area, which is the most popular approach. Description and notes can be directly processed by our system with minor modification. Among the 30 EHRs, the first ten EHRs require the CDS system to provide articles related to the patient’s diagnosis (Q1), the second ten EHRs require articles on what test should the patient receive (Q2), and last ten EHRs need CDS system to give the treatment plan articles (Q3).
We used TREC 2014 data for training the parameters in our method and used TREC 2015 and 2016 data for testing. The statistical tests were performed using Wilcoxon signed-ranks test.
Following the TREC CDS track’s setting, the evaluation metrics we used include (1) infNDCG, inferred Normalized Discounted Cumulative Gain ; (2) infAP, inferred averaged precision ; (3) P@10, precision considering only the top 10 ranked documents; and (4) MAP, mean averaged precision of all topics in the task.
To compare, we employed four basic baselines. Two low baselines, called Baseline_unigram and Baseline_MRF, only rely on the information available in the provided EHRs for generating the queries. The former one only considers the unigrams in EHRs, while the latter one considers both independent unigrams and the term dependency in medical concept clique. Between these two baselines, as shown in Table 2, Baseline_MRF significantly outperforms Baseline_unigram on 2014 and 2015 topics (p-value<0.05), but it performs significantly inferior for 2016 topics (p-value<0.05). The MRF failure on 2016 topics is due to lot of changes in the query style. In 2016 EHRs, patient’s vast disease history information appears, while in past two years, query is mostly composed by symptoms. Disease history information might lead the important symptom information less weighted. Also, the target collection size is nearly doubled. The TREC CDS track overview [3, 25] shows that the mean infNDCG performance of all participant systems drops from 20.99% in 2015 to 18.59% in 2016. Overall, unigram queries may be relatively more stable than queries generated from MRF. Trained on 2014 topic set, parameters on best infNDCG are λT= 0.4 and λU=0.6, with smoothing parameter being 2000.
The rest two baselines employed pseudo relevance feedback to act as high baselines. Relevance Feedback Model 3 (RM3) is a popular, stable, and effective pseudo relevance feedback model . These two higher baselines, Baseline_ungiram_RM3 and Baseline_MRF_RM3, provide direct comparison to our proposed query expansion methods. Trained on 2014 topic set, best parameters are extracting most informative 3 words from top 5 documents, with smoothing parameter being 500. Again, as shown in Table 2, both Baseline_ungiram_RM3 and Baseline_MRF_ RM3 significantly improve over Baseline_unigram and Baseline _MRF (p-value<0.05), respectively, on TREC CDS 2014 and 2015 topics, which indicates that they are indeed higher baselines. However, their performance improves over the two non-PRF baselines on TREC CDS 2016 data, but the improvement is not significant.
4.3 How well can predicted diseases improve CDS system performance?
In this study, for each topic, we have three predicted diseases: one from Wiki-DP, one from SMDB-DP, and the last one from the fusion of two disease ranking lists. From Table 2, we find that all three kinds of diagnosis predictions significantly improve the retrieval performance in 2014, 2015, and 2016 topics (p-value<0.05). Furthermore, Wiki-DP and fused predicted diseases even outperform the RM3 model in 2014 and 2015.
In the results of 2016 CDS, performance is much lower than 2014 and 2015. As stated above, it might be because the topics in 2016 are quite different from past years, with disease history information being introduced and the target collection size is nearly doubled.
4.3.1 The effectiveness of having Wiki-DP
In examining the effectiveness of having our diagnosis prediction algorithm using Wikipedia, we compare its performance with the baselines and some state-of-the-art systems. As shown in Table 2, our two wiki runs (Unigram_ Wiki and MRF_wiki) outperform their corresponding low baselines (Baseline_unigram and Baseline_MRF) as well as their corresponding high baselines (Baseline_ungiram_ RM3 and Baseline_MRF_RM3) in all three TREC datasets (2014, 2015, and 2016). Statistical tests show that all improvements are significant.
The last three rows in Table 2 show the best performed system among all TREC participants in each year. The performance values presented in Table 2 are based on their published work notes. We can see that, for 2014 and 2015, our two wiki runs achieve much higher performance than these best systems. Even for TREC 2016, our methods are very close to the state-of-the-art runs. This validates the effectiveness of having Wiki-DP.
4.3.2 The effectiveness of having SMDB-DP
The improvement obtained from using SMDB-DP alone is not as good as wiki-DP, but Unigram_SMDB and MRF_ SMDB still significantly outperforms Baseline_unigram and Baseline_MRF in 2014 and 2015 (p-value<0.001). However, Table 2 shows that RM3 is much better than SMDB-DP (p-value<0.001). This indicates that the concept-level diagnosis prediction is harder than word-level prediction.
We think that there might be several reasons. First, the medical concepts are not precisely extracted (above mentioned 75% accuracy), while word-level retrieval (wiki-DP) do not have this problem. Second, through MetaMap, one disease name might be identified with several concepts, for example, “Hypotension” has two CUIs, C0020649 and C3163620. In this work, we simply use the MetaMap’s top recommendation, but maybe other concepts also work or even better. In addition, since SMDB-DP has a very bad performance in 2016, it can be inferred that the vast history diseases severely affect the SMDB-DP.
4.3.3 The Combined Effectiveness of Wiki-DP and SMDB-DP
We conducted two runs on top of unigram query and MRF query with the expansion of fused results from the diagnoses from both Wiki-DP and SMDB-DP. As introduced in Section 3.4, such fusion is heavily relying on the Wiki-DP’s performance. As shown in Table 2, runs Unigram_ Fusion and MRF_Fusion have quite similar performance with Unigram_Wiki and MRF_Wiki. Although in 2015 CDS, Unigram_Fusion significantly outperforms Unigram_Wiki on infNDCG (29.32% vs 26.67%), it is probably because the parameters were trained on 2014 CDS. If it were trained on 2015 CDS data, Unigram_Wiki can get infNDCG at 29.43%, basically it is the same with Unigram_Fusion.
Among the 90 topics tested in our experiments, the results from the fused diagnosis differ from the Wiki-DP predictions by about 30% (10 different predictions in 2014, 7 in 2015, and 12 in 2016). Figure 3 shows the infNDCG performance of the 30 topics on 2014 CDS data. The fused predictions of three topics further improve the retrieval performance by a large degree, whereas the fused predictions on two topics generated inferior results against the Wiki only approach. This indicates that SMDB-DP can work as a supplementary module for the Wiki-DP, but the fused data are not promised to be always correct.
4.4 How accurate is the predicted diagnosis?
Although the experiments presented in Section 4.3 show the significant improvement contributed by the predicted diagnosis in helping biomedical retrieval, intrinsically, it is hard to identify whether the predicted diagnosis is correct or not. First, this is because we do not have ground-truth diagnosis for 70 of the 90 topics. Second, even among the 20 topics that 2015 CDS provides the correct diagnosis, it is still hard to judge the correctness of our predicted disease. For example, the TREC provided diagnosis for topic 15 is “Paroxysmal Atrial fibrillation”, and our predicated disease is “Atrial fibrillation”. This partially matched prediction can improve infNDCG by 0.16, but it is not exactly the same as true diagnosis. Therefore, in this paper, we define the usefulness of the predicted disease rather than the correctness, and state that only when a prediction can improve the topic retrieval performance by at least 1.00% on infNDCG, we state that the prediction is useful.
As shown in Table 3, Wiki-DP generated the highest portion of useful predictions, with mean portion of usefulness prediction to be 65.56%, and gave relatively robust performance across the three years’ topic sets. Figure 4 further shows the predicated diseases by Wiki-DP. SMDB-DP generated the lowest portion of useful predictions (43.33% mean value), indicating the difficulty of concept-level diagnosis prediction.
Through the experiment, we have demonstrated that our method of utilizing online open knowledge bases for diagnosis prediction to improve the medical literature retrieval can significantly improve the performance and reach to the comparable level of the state-of-the-art methods. In this section, we want to review the methods in more detail in terms of the places where it fails and the comparison with some existing approaches.
5.1 Further Analysis of Diagnosis Prediction
As shown in the results, most predicted diagnoses made by our methods are correct. Even when the retrieval from one knowledge base fails, the results from the other one can often help to recover to the correct prediction. For example, with the initial query extracted from Figure 1, “Epiglottitis” is ranked 8th in the results obtained through SemMedDB knowledge base, but it is the first in the list from Wikipedia. This helps to make it the correct prediction. Another disease appearing in both results is “Retropharyngeal abscess”, which ranks 7th in the SemMedDB results, but it is 14th in the Wikipedia list. So, drawing evidence from two sources does make our diagnosis prediction methods more robust. The results of our experiments show that, in general, the results from Wikipedia are usually more reliable, but a confirmation from the SemMedDB makes the predictions even more accurate.
However, our diagnosis prediction method does make errors. There are two types of causes to the errors in Wiki-DP. The first one is related to the insufficient information in Wikipedia data. Typical scenarios include that the correct diseases do not have sufficient content in their Wiki pages or the terms in the query to search in Wikipedia fail to distinguish the correct diseases from irrelevant diseases. For example, Topic 25 in 2015 CDS should have a correct associated disease “Osteomyelitis”, but the Wiki page of “Osteomyelitis” contains no symptom information at all. In the meantime, the Wiki page of “Langerhans-cell histiocytosis” contains some symptoms mentioned in the given EHRs. This causes the prediction method to wrongly select “Langerhans-cell histiocytosis” rather than “Osteomyelitis”.
The second type of cause is related to the limitation of current retrieval mechanism, particularly the handling of negation. For example, the EHRs associated with Topic 12 in 2014 CDS show that the patient has a symptom of “weight gain”, and the Wiki page of “Anorexia nervosa” talks about patients wanting to “prevent weight gain”, “fear weight gain”, or “avoid weight gain”. Because of lacking processing negation in collection text, our method wrongly ranks “Anorexia nervosa” as the most plausible disease for the patient.
To combat these problems, we need to enlarge our external knowledge bases to include more publicly available online resources. For example, there are published medical concept relationship datasets, such as MayoClinic6 or WebMD7. This will be a future work to explore.
In terms of SMDB-DP, there are three causes to the errors. First, the medical concepts in SemMedDB corpus are not precisely extracted (75% accuracy ), while Wiki-DP does not have this procedure and information loss. Second, through MetaMap, one disease name might be identified with several concepts. For example, MetaMap will map “Hypotension” into two medical concepts, as shown in Figure 5. In this work, we simply use the top identification from MetaMap, but other concepts should also be considered. Third, there are many medical concepts extracted from query not existing in SemMedDB. For example, “weight loss” is a symptom appearing in 4 topics, but SemMedDB do not have such a symptom in extracted concepts, making the information need not fully provided to the SMDB-DP module. In conclusion, concept-level diagnosis prediction is harder and more complicated than the word level.
5.2 Further Comparison to the State-of-the-Art Biomedical Retrieval Systems
Although our methods only reach the comparable performance with the state-of-the-art algorithms, our retrieval model is much simpler. For example, the best performed system in the 2014 CDS task was developed by Choi et al. . They used pseudo relevance feedback (PRF) to obtain most frequently mentioned MESH terms from the top-ranked articles and expand their queries with such terms. This increased their algorithm’s performance from 19.21% to 22.24% on infNDCG. Then their method used classification method to re-rank the results to achieve their best results, 26.74%. In contrast, our method only performs query expansion with predicted diagnosis, which increases in infNDCG from 21.88% to 28.44%. We believe that our approach is simpler in retrieval model. At the same time, we regard the CDS task as a concept-based information retrieval task, in which our query can recognize the concept associated with a disease but is not limited by the text expression in the document.
The best run in 2015 CDS task was submitted by Balaneshin-kordan et al. . They explored a lot on medical concept detection and selection and expanded their queries with the most important unigrams in PRF documents. They attributed their success to the MRF model and Parameterized Query Expansion (PQE). However, we cannot tell them apart since they did not publish the intermediate performance.
The best run in 2016 CDS task was submitted by Gurulingappa et al. . They first expanded the original query with extracted UMLS medical concepts from EHRs and with the most important words in PRF. Then, they measured the document similarity based on word embeddings and combined these features with a learning-to-rank model which improved the infNDCG from 22.61% to 24.93%. This method indicates that the word embedding can help search the relevant documents that use different words but keep the same relevant information.
In summary, our method takes a different route to the existing state-of-the-art methods. It is possible that our method can be combined with these state-of-the-art approaches or even more advanced retrieval methods. Under such situation, the retrieval improvement can be even larger.
In this paper, we target to enhance the current CDS systems with public knowledge bases, Wikipedia and SemMedDB. To be specific, a word-level and a concept-level diagnosis prediction methods are proposed to automatically find the disease of the patient, which are used to perform query expansion in medical text retrieval. This idea has been proven to be effective by the significantly improved retrieval performance using our methods through the validation on TREC CDS track data of 2014, 2015, and 2016. Our disease prediction accuracy can reach 65.56%. In the future, we will incorporate the word embedding techniques to enhance the diagnosis prediction methods.
This work was partially supported by Wuhan University’s independent research project (Humanities and Social Sciences) “Human-Computer Interaction and Collaboration Team” (Whu2016020).
Ely, J. W. et al. (2000). A taxonomy of generic clinical questions: classification study. BMJ 321.7258, 429-432. Google Scholar
Simpson, M. S., Voorhees, E. M. & Hersh, W. Overview of the trec 2014 clinical decision support track. Lister Hill National Center for Biomedical Communications Bethesda MD. Google Scholar
Roberts, K. et al. (2015). Overview of the TREC 2015 Clinical Decision Support Track. Proceedings of the 2015 Text Retrieval Conference. Google Scholar
Zighelnic, L. & Kurland, O. (2008). Query-drift prevention for robust query expansion. Proceedings of the 31st annual international ACM SIGIR Conference on Research and Development in Information Retrieval. ACM. Google Scholar
Carpineto, C. & Romano, G. (2012). A survey of automatic query expansion in information retrieval. ACM Computing Surveys (CSUR) 44(1), 1. Google Scholar
Choi, S. & Choi, J. (2014). SNUMedinfo at TREC CDS track 2014: Medical case-based retrieval task. Seoul National University, Republic of Korea. Google Scholar
Oh, H-S. and Jung, Y. (2015). Cluster-based query expansion using external collections in medical information retrieval. Journal of Biomedical Informatics 58, 70-79. PubMedCrossrefWeb of ScienceGoogle Scholar
Balaneshin-kordan, S., Kotov, A. & Xisto, R. (2015). WSU-IR at TREC 2015 Clinical Decision Support Track: Joint Weighting of Explicit and Latent Medical Query Concepts from Diverse Sources, Proceedings of the 2015 Text Retrieval Conference. Google Scholar
Xie, Z., Xia, Y. & Zhou, Q. (2015). Incorporating Semantic Knowledge with MRF Term Dependency Model in Medical Document Retrieval, National CCF Conference on Natural Language Processing and Chinese Computing: Springer International Publishing. Google Scholar
Song, Y. et al. (2015). ECNU at 2015 CDS Track: Two Re-ranking Methods in Medical Information Retrieval, Proceedings of the 2015 Text Retrieval Conference. Google Scholar
Bodenreider, O. (2004).The unified medical language system (UMLS): integrating biomedical terminology, Nucleic Acids Research 32.suppl 1, D267-D270. Google Scholar
Aronson, A. R. (2001). Effective mapping of biomedical text to the UMLS Metathesaurus: the MetaMap program, Proceedings of the AMIA Symposium. American Medical Informatics Association. Google Scholar
Voorhees, E. M. (2014). The effect of sampling strategy on inferred measures, Proceedings of the 37th international ACM SIGIR Conference on Research & Development in Information Retrieval. ACM. Google Scholar
Strohman, T. et al. (2005). Indri: A language model-based search engine for complex queries.” Proceedings of the International Conference on Intelligent Analysis. Vol. 2. No. 6. Google Scholar
Isola, R., Carvalho, R. & Tripathy, A. K. (2012). Knowledge Discovery in Medical Systems Using Differential Diagnosis, LAMSTAR, and $ k $-NN, IEEE Transactions on Information Technology in Biomedicine 16(6), 1287-1295. Web of ScienceCrossrefGoogle Scholar
Gomathi, P. & Nithya, NS. Medical Disease Diagnosis Using Structuring Text, International Journal of Computer Science & Engineering Technology 1(5): 591-594. Google Scholar
Liu, C. et al. (2016). Augmented LSTM Framework to Construct Medical Self-diagnosis Android, Data Mining (ICDM), 2016 IEEE 16th International Conference on. IEEE. Google Scholar
Koopman, B., Zuccon, G., Nguyen, A., Bergheim, A. & Grayson, N. (2015). Automatic ICD-10 classification of cancers from free-text death certificates. International Journal of Medical Informatics 84(11):956-65. CrossrefPubMedWeb of ScienceGoogle Scholar
Metzler, D. & Croft, W. B. (2005). A Markov random field model for term dependencies, Proceedings of the 28th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM. Google Scholar
Gurulingappa, H. et al. (2016). Semi-Supervised Information Retrieval System for Clinical Decision Support, Proceedings of the 2016 Text Retrieval Conference. Google Scholar
Roberts, K. et al. (2016). Overview of the TREC 2016 Clinical Decision Support Track, Proceedings of the 2016 Text Retrieval Conference. Google Scholar
Nie, L., et al. (2015). Disease inference from health-related questions via sparse deep learning, IEEE Transactions on Knowledge and Data Engineering 27(8), 2107-2119. CrossrefWeb of ScienceGoogle Scholar
Lv, Y. & Zhai, C. (2009). A comparative study of methods for estimating query language models with pseudo feedback, Proceedings of the 18th ACM Conference on Information and Knowledge Management. ACM. Google Scholar
PubMed is a medical literature corpus, comprising more than 27 million citations. Available at https://www.ncbi.nlm.nih.gov/pubmed/.
SemMedDB is accessible in https://skr3.nlm.nih.gov/SemMedDB/dbinfo.html
About the article
Published Online: 2017-09-29
Citation Information: Data and Information Management, Volume 1, Issue 1, Pages 49–60, ISSN (Online) 2543-9251, DOI: https://doi.org/10.1515/dim-2017-0005.
© 2017 Danchen Zhang, Daqing He. This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 License. BY-NC-ND 3.0