Injecting Domain Knowledge in Electronic Medical Records to Improve Hospitalization Prediction

Electronic medical records (EMRs) contain essential information about the different symptomatic episodes a patient goes through. They have the potential to improve patient well-being and are therefore a potentially valuable source to artificial intelligence approaches. However, the linguistic variety as well as the tacit knowledge in EMRs can impair the prognosis of a machine learning algorithm.

In this paper, we extract ontological knowledge from text fields contained in EMRs and evaluate the benefit when predicting hospitalization. Our study uses a dataset extracted from the PRIMEGE PACA relational database [10] which contains more than 350,000 consultations in French by 16 general practitioners (Table 1). In this database, text descriptions written by general practitioners are available with international classification codes of prescribed drugs, pathologies and reasons for consultations, as well as the numerical values of the different medical examination results obtained by a patient.

Our initial observation was that the knowledge available in a database such as PRIMEGE remains limited to the specificities of each patient and in particular that the texts found in there are based on a certain amount of implicit information known to medical experts. Moreover, the level of detail of the information contained in a patient’s file is variable. Therefore, a machine learning algorithm exploiting solely the information at its disposal in an EMR will not be able to exploit this specific knowledge implicit in the documents it analyzes or, at best, it will have to relearn this knowledge by itself, possibly in an incomplete and costly way.

In that context, our main research question is: Can ontological augmentations of the features improve the prediction of the occurrence of an event? In our case study, we aim to predict a patient’s hospitalization using knowledge from different knowledge graphs in the medical field. In this paper, we focus on the following sub-questions:

• How to integrate domain knowledge into a vector representation used by a machine learning algorithm?

• Is the addition of domain knowledge improving the prediction of a patient’s hospitalization?

• Which domain knowledge combined with which machine learning methods provide the best prediction of a patient’s hospitalization?

To answer these questions, we first survey the related work (Sect. 2) and position our contribution. We then introduce the proposed method for semantic annotation and knowledge extraction from texts and specify how ontological knowledge is injected upstream into the vector representation of EMRs (Sect. 3). Then, we present the experimental protocol and discuss the results obtained (Sect. 4). Finally, we conclude and provide our perspectives for this study (Sect. 5).

In [12], the authors are focused on finding rules for the activities of daily living of cancer patients on the SEER-MHOS (Surveillance, Epidemiology, and End Results - Medicare Health Outcomes Survey) and they showed an improvement in the coverage of the inferred rules and their interpretations by adding ‘IS-A’ knowledge from the Unified Medical Language System (UMLS). They extract the complete sub-hierarchy of kinship and co-hyponymous concepts. Although their purpose is different from ours, their use of the OWL representation of UMLS with a machine learning algorithm improves the coverage of the identified rules. However, their work is based solely on ‘IS-A’ relationships without exploring the contributions of other kinds of relationships and they do not study the impact of this augmentation on different machine learning approaches: they used the AQ21 and the extension of this algorithm AQ21-OG to compare.

In [5], the authors established a neural network with graph-based attention model that exploits ancestors extracted from the OWL-SKOS representations of ICD Disease, Clinical Classifications Software (CCS) and Systematized Nomenclature of Medicine Clinical Terms (SNOMED-CT). In order to exploit the hierarchical concepts of these knowledge graphs in their attention mechanism, the graphs are transformed using the embedding obtained with Glove [15]. The results show that such a model better performs when identifying a pathology rarely observed in a training dataset than a recurrent neural network and it also better generalizes when confronted with less data in the training set. Again, this work also does not exploit other kinds of relationships, while we compare the impact of different kinds and sources of knowledge.

In [16] the authors extract knowledge from the dataset of [13] and structure it with an ontology developed for this purpose, then they automatically deduce new class expressions, with the objective of extracting their attributes to recognize activities of daily living using machine learning algorithms. The authors highlight better accuracy and results than with traditional approaches, regardless of the machine learning algorithm on which this task has been addressed (up to 1.9% on average). Although they exploit solely the ontology developed specifically for the purpose of discovering new rules, without trying to exploit other knowledge sources where a mapping could have been done, their study shows the value of structured knowledge in classification tasks. We intend here to study the same kind of impact but with different knowledge sources and for the task of predicting hospitalization.

3.1 Extraction of Ontological Knowledge from EMRs

Our study aims to analyze and compare the impact of knowledge from different sources, whether separately incorporated or combined, on the vector representation of patients’ medical records to predict hospitalization. To extract domain knowledge underlying terms used in text descriptions written by general practitioners, we search the texts for medical entities and link them to the concepts to which they correspond in Wikidata, DBpedia and health sector specific knowledge graphs such as those related to drugs. Wikidata and DBpedia were chosen because general concepts can only be identified with general repositories. In this section, we describe how these extractions are performed but we do not focus on this step as it is only a means to an end for our study and could be replaced by other approaches.

Knowledge Extraction Based on DBpedia. To detect in an EMR concepts from the medical domain present in DBpedia, we used the semantic annotator DBpedia Spotlight [7]. Together with the domain experts, we carried out a manual analysis of the named entities detected on a sample of approximately 40 consultations with complete information and determined 14 SKOS top concepts designating medical subjects relevant to the prediction of hospitalization, as they relate to severe pathologies (Table 2).

For each EMR to model, from the list of resources identified by DBpedia Spotlight, we query the access point of the French-speaking chapter of DBpedia^{Footnote 1} to determine if these resources have as subject (property dcterms:subject) one or more of the 14 selected concepts.

In order to improve DBpedia Spotlight’s detection capabilities, words or abbreviated expressions within medical reports are added to text fields using a symbolic approach, with rules and dictionaries (e.g., the abbreviation “ic” which means “heart failure” is not recognized by DBpedia Spotlight, but is thus correctly identified through this symbolic approach).

In the rest of the article, the \(+s\) notation refers to an approach using the enrichment of representations with concepts from DBpedia according to the method previously described. The \(+s*\) notation refers to an approach that does not exploit all text fields and extracts concepts from text fields related to the patient’s personal history, allergies, environmental factors, current health problems, reasons for consultations, diagnoses, medications, care procedures, reasons for prescribing medications and physician observations. The \(+s*\) approach focuses on the patient’s own record, not on his family history and past problems. Note that the symptom field is used by doctors for various purposes and this is the reason why we excluded this field in the DBpedia concept extraction procedure for this approach.

Knowledge Extraction Based on Wikidata. Wikidata^{Footnote 2} is an open knowledge base that centralizes data from various projects of the Wikimedia Foundation. Its coverage on some domains differs from that of DBpedia. We extracted drug-related knowledge by querying Wikidata’s endpoint.^{Footnote 3} More precisely, we identified three properties of drugs relevant to the prediction of hospitalization: ‘subject has role’ (property wdt:P2868), ‘significant drug interaction’ (property wdt:P2175), and ‘medical condition treated’ (property wdt:P769).

In Wikidata, we identify the drugs present in EMRs using the ATC code (property wdt:P267) of the drugs present in the PRIMEGE database. The CUI UMLS codes (property wdt:P2892) and CUI RxNorm (property wdt:P3345) have been recovered using medical domain specific ontologies. Indeed, the codes from these three representations are not necessarily all present to identify a drug in Wikidata, but at least one of them allows us to find the resource about a given drug. From the URI of a drug, we extract property-concept pairs related to the drugs for three selected properties (e.g., ‘Pethidine’ is a narcotic, ‘Meprobamate’ cures headache, ‘Atazanavir’ interacts with ‘Rabeprazole’).

In the rest of the article, the notation \(+wa\) refers to an approach using the enrichment of our representations with the property ‘acts as such’, \(+wm\) indicates the usage of the property ‘treated disease’ and \(+wi\) of the property ‘drugs interacts with’.

Knowledge Extraction Based on Domain Specific Ontologies. We were interested in the impact of contributions from domain specific knowledge graphs especially for text fields containing international drug codes from the Anatomical, Therapeutic and Chemical (ATC) classification and codes related to the reasons for consulting a general practitioner with the International Classification of Primary Care (CISP-2). We thus extracted knowledge based on three OWL representations specific to the medical domain: ATC^{Footnote 4}, NDF-RT^{Footnote 5} and CISP2.^{Footnote 6} The choice of OWL-SKOS representations of CISP2 and ATC in our study comes from the fact that the PRIMEGE database adopts these nomenclatures, while the OWL representation of NDF-RT provides additional knowledge on interactions between drugs, diseases, mental and physical states.

We extracted from the ATC OWL-SKOS representation the labels of the superclasses of the drugs listed in the PRIMEGE database, using the properties rdfs:subClassOf and member_of on different depth levels thanks to SPARQL 1.1 queries with property paths^{Footnote 7} (e.g. ‘meprednisone’ (ATC code: H02AB15) has as superclass ‘Glucocorticoids, Systemic’ (ATC code: H02AB) which itself has as superclass ‘CORTICOSTEROIDS FOR SYSTEMIC USE, PLAIN’ (ATC code: H02)).

Similarly, we extracted from the OWL-SKOS representation of CISP2 the labels of the superclasses with property rdfs:subClassOf, however, given the limited depth of this representation, it is only possible to extract one superclass per diagnosed health problem or identified care procedure (e.g., ‘Symptom and complaints’ (CISP-2 code: H05) has for superclass ‘Ear’ (CISP-2 code: H)).

In the OWL representation of NDF-RT, we selected three drug properties relevant to the prediction of hospitalization: ‘may_treat’ property (e.g. ‘Tahor’, whose main molecule is ‘Atorvastatin’ (ATC code: C10AA05) can cure ‘Hyperlipoproteinemias’ (Hyperlipidemia), ‘CI_with’ (e.g. ‘Tahor’ is contraindicated in ‘Pregnancy’) and ‘may_prevent’ (e.g. ‘Tahor’ can prevent ‘Coronary Artery Disease’)). A dimension in our DME vector representation will be a property-value pair. Here is an example RDF description of the drug Tahor:

In the rest of the article, the notation \(+c\) refers to an approach using the enrichment of vector representations with ATC and the number attached specifies the different depth levels used (e.g., \(+c_{1-3}\) indicates that 3 superclass depth levels are integrated in the same vector representation). \(+t\) indicates the enrichment of vector representations with CISP2. \(+d\) indicates the enrichment of vector representations with NDF-RT, followed in indices by CI if property ‘CI_with’ is used, prevent if property ‘may_prevent’ is used, treat if property ‘may_treat’ is used. For example, \(+d_{CI,prevent,treat}\) refers to the case where these three properties are used together in the same vector representation of DMEs.

3.2 Integrating Ontological Knowledge in a Vector Representation

It is crucial when using a domain-specific corpus to generate its own representation, since many terms may be omitted in a general representation or an ambiguous notion may be applied to a term when it has a very precise definition in a given sector. We opted for a model using a bag-of-words representation (BOW) for different reasons: (i) the main information from textual documents is extracted without requiring a large corpus; (ii) the attributes are not transformed, which makes it possible to identify which terms contribute to the distinction of patients to hospitalize or not, even if this implies to manipulate very large vector spaces; (iii) the integration of heterogeneous data is facilitated since it is sufficient to concatenate other attributes to this model without removing the meaning of the terms previously represented in this way.

Just like in the structure of the PRIMEGE database, some textual data must be distinguishable from each other when switching to the vector representation of EMRs, e.g. a patient’s personal history and family history. To do this, we have introduced provenance prefixes during the creation of the bag-of-words to trace the contribution of the different fields.

Concepts from knowledge graphs are considered as a token in a textual message. When a concept is identified in a patient’s medical record, it is added to a concept vector. This attribute will have as value the number of occurrences of this concept within the patient’s health record (e.g., the concepts ‘Organ Failure’ and ‘Medical emergencies’ are identified for ‘pancréatite aiguë’, acute pancreatitis, and the value for these attributes in our concept vector will be equal to 1).

Similarly, if a property-concept pair is extracted from a knowledge graph, it is added to the concept vector. For example, in vectors exploiting the NDF-RT (vector representation d), we find the couple consisting of CI_with as a property - contraindicated with- and the name of a pathology or condition, for example ‘Pregnancy’.

Workflow diagram to generate vector representations integrating ontological knowledge alongside with textual information.

Full size image

Concept vectors generated for two EMRs with the bag-of-words approach under the \(+s\) configuration. The translation and correction of the texts are (a) for patient 1: “predom[inates] on the left, venous or cardiac insuf[ficiency], no evidence of phlebitis, does not want to wear compression stockings and does not want to increase the lasix”. and (b) for patient 2: “In vitro fertilization procedure, embryo transfer last Saturday, did ovarian hyperstimulation, cyst rupture, asthenia, abdominal [pain], [pain] on palpation ++, will see a gyneco[logist] next week [for] a beta HCG, echo check-up”.

Full size image

Let \(V^i=\{w_{1}^{i},w_{2}^{i},..., w_{n}^{i}\}\) be the bag-of-words obtained from the textual data in the EMR of the \(i^{th}\) patient. Let \(C^i=\{c_{1}^{i},c_{2}^{i},..., c_{n}^{i}\}\) be the bag of concepts for the \(i^{th}\) patient resulting from the extraction of concepts belonging to knowledge graphs after analysis of his consultations from semi-structured data such as text fields listing drugs and pathologies with their related codes, and unstructured data from free texts such as observations. The different machine learning algorithms exploit the aggregation of these two vectors: \(x^{i}=V^{i}\oplus C^{i}\) (Fig. 1).

From the following sentence “prédom à gche - insuf vnse ou insuf cardiaque - pas signe de phlébite - - ne veut pas mettre de bas de contention et ne veut pas augmenter le lasilix... -” (predom[inates] on the left, venous or cardiac insuf[ficiency], no evidence of phlebitis, does not want to wear compression stockings and does not want to increase the lasix...) the expression ‘insuf cardiaque’, meaning ‘heart failure’, refers to two concepts listed in Table 2: ‘Organ failure’ and ‘Cardiovascular disease’, these concepts were retrieved by the property dcterms:subject from DBpedia. The concept (occurrence) vector that represents the patient’s EMR will therefore have a value of 1 for the attributes representing the concepts ‘Organ Failure’ and ‘Cardiovascular Disease’ (Fig. 2).

As for the exploitation of NDF-RT, let us consider again the example description of the drug Tahor introduced in Sect. 3.1. It can be used to enrich the vector representation of the EMR of patients under Tahor as detailed in Table 3.

4.1 Dataset and Protocol

We tested and evaluated our approach for enriching the vector representation of EMRs with ontological knowledge on a balanced dataset \(DS_B\) containing data on 714 patients hospitalized and 732 patients not hospitalized. When the observation field is filled in by general practitioners, it can go from 50 characters to 300 characters on average. The best filled in fields concern prescribed drugs and reasons for consultations, then come the antecedents and active problems.

Since we use non-sequential machine learning algorithms to assess the enrichment of ontological knowledge, we had to aggregate all patients’ consultations in order to overcome the temporal dimension inherent in symptomatic episodes occurring during a patient’s lifetime. Thus, all consultations occurring before hospitalization are aggregated into a vector representation of the patient’s medical file. For patients who have not been hospitalized, all their consultations are aggregated. Thus, the text fields previously described (Table 1) are transformed into vectors.

We evaluated the vector representations by nested cross-validation [3], with an external loop with a K fixed at 10 and for the internal loop a K fixed at 3 with exploration of hyperparameters by random search [1] over 150 iterations.

The different experiments were conducted on an HP EliteBook 840 G2, 2.6 GHz, 16 GB RAM with a virtual environment under Python 3.6.3 as well as a Precision Tower 5810, 3.7 GHz, 64 GB RAM with a virtual environment under Python 3.5.4. The creation of vector representations was done on the HP EliteBook and on this same machine were deployed DBpedia Spotlight as well as domain-specific ontologies with Corese Semantic Web Factory [6],^{Footnote 8} a software platform for the Semantic Web. It implements RDF, RDFS, SPARQL 1.1 Query & Update, OWL RL.

4.2 Selected Machine Learning Algorithms

We performed the hospitalization prediction task with different state-of-the-art algorithms available in the Scikit-Learn library [14]:

• SVC: Support vector machine (SVC stands for ‘Support Vector Classification’) whose implementation is based on the libsvm implementation [4]. The regularization coefficient C, the kernel used by the algorithm and the gamma coefficient of the kernel were determined by nested cross-validation.

• RF: The random forest algorithm [2]. The number of trees in the forest, the maximum tree depth, the minimum number of samples required to divide an internal node, the minimum number of samples required to be at a leaf node and the maximum number of leaf nodes were determined by nested cross-validation.

• Log: The algorithm of logistic regression [11]. The regularization coefficient C and the norm used in the penalization were determined by nested cross-validation.

We opted for a bag-of-words model and the above cited machine learning algorithms since it is possible to provide a native interpretation of the decision of these algorithms, thus allowing the physician to specify the reasons for hospitalizing a patient with the factors on which he can operate to prevent this event from occurring. Moreover, logistic regression and random forest algorithms are widely used in order to predict risk factors in EHR [9]. Finally, the limited size of our dataset excluded neural networks approaches.

4.3 Results

In order to assess the value of ontological knowledge, we evaluated the performance of the machine learning algorithms by using the \(F_{tp,fp}\) metric [8]. Let TN be the number of negative instances correctly classified (True Negative), FP the number of negative instances incorrectly classified (False Positive), FN the number of positive instances incorrectly classified (False Negative) and TP the number of positive instances correctly classified (True Positive).

Table 4 summarizes the results for each representation and method combination tested on the \(DS_B\) dataset:

• baseline: represents our basis of comparison where no ontological enrichment is made on EMR data i.e. only text data in the form of bag-of-words.

• \(+s\): refers to an enrichment with concepts from the DBpedia knowledge base.

• \(+s*\): refers to an enrichment with concepts from the DBpedia knowledge base, unlike \(+s\), not all text fields are exploited, thus, concepts from fields related to the patient’s personal history, allergies, environmental factors, current health problems, reasons for consultations, diagnoses, medications, care procedures followed, reasons for prescribing medications and physician observations are extracted.

• \(+t\): refers to an enrichment with concepts from the OWL-SKOS representation of CISP-2.

• \(+c\): refers to an enrichment with concepts from the OWL-SKOS representation of ATC, the number or number interval indicates the different hierarchical depth levels used.

• \(+wa\): refers to an enrichment with Wikidata’s ‘subject has role’ property (wdt:P2868).

• \(+wi\): refers to an enrichment with Wikidata’s ‘significant drug interaction’ property (wdt:P769).

• \(+wm\): refers to an enrichment with Wikidata’s ‘medical condition treated’ property (wdt:P2175).

• \(+d\): refers to an enrichment with concepts from the NDF-RT OWL representation, \(_\text {prevent}\) indicates the use of the may_prevent property, \(_\text {treat}\) the may_treat property and \(_\text {CI}\) the CI_with property.

4.4 Discussion

In general terms, knowledge graphs improve the detection of true positives cases, hospitalized patients correctly identified as such (Tables 5 and 6) and provide a broader knowledge of the data in patient files like the type of health problem with CISP-2 (Table 7).

Although the approach in combination with \(+s*\) does not achieve the best final results, this approach achieves the best overall performance among all the approaches tested with 0.858 under logistic regression when using 8 K folds from the KFold during the training phase (Fig. 3). It also surpasses other methods under 3 K partitions by exceeding the baseline by 0.9% and at 4 K partitions by 0.7% \(+t+s+c_2+wa+wi\) which suggests an improvement in classification results if we enrich a small dataset with attributes from the enrichment provided by knowledge graphs.

Convergence curve obtained following the training on n (x-axis) KFold partitions for different configurations of the Table 4.

Full size image

Despite the shallow OWL-SKOS representation of CISP2, the \(+t\) configuration is sufficient to improve a patient’s hospitalization predictions, if we compare its results to those of the baseline. Surprisingly enough, a second level of super class hierarchy with \(+c_2\) from the ATC OWL-SKOS representation provides better results while only one level of hierarchy with \(+c_1\), seems to have a negative impact on them, this can be explained by the fact that the introduction of a large number of attributes ultimately provides little information, unlike the second level of hierarchy.

However the results show that applying DBpedia expansion to fields indirectly related to the patient’s condition, such as family history, can lead machine learning algorithms to draw wrong conclusions even if prefixes have been added to distinguish provenance fields. The text field related to symptoms has been poorly filled in by doctors – as it was an ‘observation’ field – and the majority of the remarks thus detected by DBpedia Spotlight are mostly false alerts. Moreover, the qualitative analysis of the results showed cases involving negation (‘pas de SC d’insuffisance cardiaque’, meaning ‘no symptom of heart failure’) and poor consideration of several terms (‘brûlures mictionnelles’, related to bladder infection, are associated with ‘Brûlure’, a burn, which, therefore, has as subject the concept ‘Urgence médicale’, a medical emergency). Both cases are current limitations of our approach and we consider for our future work the need for handling negation and complex expressions.

In this paper, we have presented a method for combining knowledge from knowledge graphs, whether specialized or generalist, and textual information to predict the hospitalization of patients. To do this, we generated different vector representations coupling concept vectors and bag-of-words and then evaluated their performance for prediction with different machine learning algorithms.

In the short term, we plan to identify additional concepts involved in predicting hospitalization of patients and to evaluate the impact of additional domain specific knowledge, as we focused mainly on drugs in our study. We also intend to propose an approach to automatically extract candidate medical concepts from DBpedia. Finally we plan to improve the coupling of semantic relationships and textual data in our vector representation, and support the detection of negation and complex expressions in texts.