Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.
Recruitment of clinical research participants is routinely disappointing with traditional methods failing to identify up to 60% of possible participants.1 2 Substantial institutional and departmental expense is incurred and little scientific benefit is gained by low-enrolling studies, which made up 31% of the studies at one institution over a single year.3 Evidence indicates that dramatic increases, up to fourfold, in recruitment are possible with automated recruitment.4 5 Such approaches are scalable in research settings—some research institutions have linked, with proper privacy safeguards in place, electronic medical records (EMRs) data together with genotype data for discovery in large-scale databases and virtual cohorts.6 EMR analysis has been suggested as a useful means of measuring outcomes and defining disorder subpopulations.7
Research inclusion criteria in psychiatry often use diagnosis. Structured diagnosis codes are sometimes available in EMR clinical notes, but are frequently missing. Procedures such as natural language processing (NLP) and machine learning (ML) methods have been used to extract clinical information from EMRs’ unstructured text. The eMERGE group has used NLP extensively, with improved accuracy of their phenotyping algorithms8—examples include determining colorectal cancer screening status9 and diagnosing rheumatoid arthritis.10 NLP methods have also been applied to EMRs to boost the efficiency of manual chart abstraction for breast cancer recurrence with 92% sensitivity and 96% sensitivity.11 More recently, NLP has been used to identify adverse drug events including extrapyramidal side effects in psychiatric patients12 and to phenotype children at risk for Kawasaki disease in emergency department notes.13 In another investigation, ML classification algorithms were used to identify rheumatoid arthritis patients with coronary artery disease—NLP was used to detect features in clinical notes and outperformed features selected by experts.14
In recent years, mental health researchers in South London and Maudsley NHS Trust have begun using EMRs for research recruitment.15 16 For phenotyping, a small number of studies have focused on extracting depression diagnoses from unstructured EMR text. Early work on diabetes outpatient records compared diagnosis by coding versus by NLP—NLP improved detection of depression diagnosis by almost a third.17 Researchers developed and tested NLP in patients with a billing code of major depressive disorder to characterise symptom remission and treatment resistance, and found that adding NLP resulted in higher area under receiver operating characteristic curve than billing data only (0.85–0.88 vs 0.54–0.55) for classification of mood state.18 NLP has been used for categorisation of publicly available Twitter data into several mental health diagnoses, including depression and bipolar disorder.19 A later publication identified patients with depression from free-text discharge summaries: a combination of NLP and ML algorithms was used, with the best performance coming from Medical Text Extraction, Reasoning and Mapping System’s20 knowledge-based decision tree method, yielding an F-measure of 89.6%.21 To summarise the rationale for extracting diagnosis inclusion criteria from unstructured EMR using NLP and ML, it is known that research recruitment supported by automation is more successful; further, that NLP and ML can be useful for information extraction from unstructured text notes, and that such methods have been applied with some degree of success to depression-related phenotypes.
Since structured diagnosis codes had limited availability in our EMR, we used NLP and ML on EMR notes data to extract our diagnostic inclusion criteria, in this case Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV depression diagnoses, to support recruitment for a cognitive-genomic study of youth depression. This article summarises the NLP and ML processes and results. The core purpose of this report is to present a model that identifies youth with a depression diagnosis and without specific exclusion comorbidities—a model evaluated via cross-validation and an independent test data set, based on deep neural networks.
Clinical documents commonly contain sensitive information about individuals; accordingly, in this Research Ethics Board-approved study, we deidentified the corpus to remove personal identifying information (PII). For this task, we created a suite of programs that made use of the freely available Perl-based software package De-id V.1.1.22 With these programs, we performed the following tasks:
Inserted the necessary text tags at the beginning and end of each document so that it would be recognised by De-id
Converted all the documents to. txt files so that the format conformed to the De-id specifications
Looped the De-id algorithm over the whole document corpus to remove PII and thereby coded them
Translated the coded documents into. csv files to get them ready for the training and testing protocols of our supervised learning methods.
Clinical documents for youth psychiatric patients often contain important free-text information regarding a patient’s lifestyle, activities and clinical impressions, including diagnosis; however, often a discrete/structured diagnosis is missing. Our aim is to use NLP and ML to identify our phenotype of interest: youth patients ages 12–18 with DSM-IV defined Major Depressive Disorder or Dysthymic Disorder. Exclusion criteria included schizophrenia, bipolar disorder, autism, epilepsy, personality disorder, developmental delay and traumatic brain injury. From our EMR, we obtained a corpus consisting of 861 physician documents on 366 patients ages 12–18 years for a 6-month period, and deidentified them as noted above; the documents were predominantly progress notes, with character counts ranging from 533 to 24 803 (without spaces) and a median character count of ~4300. Almost all the child and adolescent patient population at the Centre for Addiction and Mental Health is outpatient in nature. Of the corpus, 60% of documents were on females.
This specific phenotyping effort requires a model that is capable of rejecting documents of individuals manifesting the exclusion criteria, but simultaneously requires a model capable of including suitable participants’ documents.
We used two distinct approaches for this task: (i) a brute force search method based on specific terms stored in dictionaries and (ii) an ML protocol known as neural networks.23 Both methods relied on NLP packages/methods available through the R programming language: (wordnet, RKEA, tm, SDMTools). These methods take the EMR clinical document corpus and translate it into a structure that allows a machine to efficiently compute the frequency structure of the words used in each document; the term frequencies are recorded in the Document Term Matrix (DTM) (see table 1).24 The NLP methods assure that only meaningful words are used by performing functions such as stripping grammatical articles from the text. For the neural network algorithms we present, the DTM (table 1) is the data that is being used by the brute force and neural network algorithms to find potential study participants.
The DTM records information about how often a word is encountered in a document but, in our case, it also includes information about how often it is found in the full document set. In table 1, each row represents a document from a patient, and the columns are words. We use the tf-idf computation (term frequency-inverse document frequency) which captures information about how often words show up in a document but it also adjusts for the effect of high-frequency words. A word such as ‘the’ or ‘diagnosis’ may add little information and their influence appropriately minimised. This approach allows our algorithms to focus on terms that ‘stand out’.
We used a supervised learning paradigm—we applied labels to the documents, for the algorithm to learn from, that is, suitable research participant candidates or not. To label a data set of EMR documents, two fully qualified psychiatrists (AR and JS) independently annotated 900 patient documents, which resulted in 861, after omitting 39 unclassifiable documents. Agreement between annotators was 98% based on 100 documents annotated by both psychiatrists. Of this set, there were 126 documents that were classified as belonging to patients that would meet the above inclusion criteria, and not meet any criteria for exclusion.
The brute force method attempted to identify suitable participants by looking for certain keywords that would cause the machine to either reject or accept a particular document as belonging to a patient that would make a suitable participant. The method used a positive dictionary (PD) for inclusion criteria diagnoses and a negative dictionary (ND) for exclusion criteria diagnoses, along with a subalgorithm that looked at words that come before or after the specific PD or ND words. The words in the PD would increase a score of acceptance for an EMR and words in the ND would have the opposite effect. The subalgorithm that looked at the surrounding words would decide if the words in either the PD or ND should be negated, for example, ‘it is unlikely Samantha has major depressive disorder’ (box 1, 2).
Positive dictionary: a dictionary of terms to help identify depression
Major depressive disorder
Persistent depressive disorder
Negative dictionary: terms that would indicate that someone is not suitable
Traumatic brain injury
Borderline personality disorder
Neural networks have received more attention in recent years mainly due to advancements in methodology and access to affordable powerful computation platforms. The popularity of what are known as deep neural networks stems from their ability to robustly identify images.23 Advances in the last decade have been very impressive for image classification25 in addition to NLP.26 We decided to use the deep learning paradigm (DL) because of the expected non-linear relationships that exist between the language used within the EMRs and DL’s ability to learn several representations simultaneously for distinguishing between suitable participants and not. Deep neural networks encode information to make a prediction in a way that uses several layers of information by making non-linear inferences between the variables—in this case the frequencies of used words and co-occurrences of used words. If the two groups of patients were linearly separable, then such a sophisticated method would not be necessary, and indeed, for a subset of patients this is true as some documents contain clear diagnoses. However, we are using our ability to move beyond a simple search, as was implemented in the brute force approach, via deep neural networks.
We used an R language implementation of the H2O.ai package, which includes a multilayer, feedforward deep neural network for the purpose of prediction under a supervised protocol. For more details, please refer to H2O open-source software.25 We used the 861 documents as two main data sets of 758 and 103 documents, respectively: (i) a training data set consisted of 758 documents, with 101 suitable participants and 657 unsuitable participants; and (ii) a test data set consisted of 103 documents with 25 of them belonging to suitable participants and 78 unsuitable participants. Our training phase resulted in two models that we shall refer to as DL1 and DL0: DL1 is capable of accurately identifying suitable participants but is poor at identifying unsuitable participants and DL0 has the opposite capabilities, as it is very accurate at correctly rejecting participants. Test statistics will be provided in the ‘Results’ section.
These two models were combined into a single protocol that takes patient documents as input and provides a list of patients for inclusion in our study. We shall refer to this model as DL1+0, which works by first passing a new group of patients to evaluate through DL1. The DL1+0 method will then provide a label for each patient by evaluating the corresponding document. At this stage DL1+0 will capture a good proportion of the true candidates but it will likely label many unsuitable candidates as suitable, so it then passes this new smaller list of documents through DL0, which then removes documents of patients that it deems to be unsuitable, thus ending up with a list of proposed true potential participants. See figure 1 for a synopsis of this process.
For information regarding De-id performance, please refer to Neamatullah and colleagues.22 We customised De-id for our purposes to include a larger set of proper nouns including names and regional institutional names for more optimal deidentification. The performance statistics presented here relate to individual documents, not patients.
The brute force method was capable of performing well on some data sets but it did not generalise well. On some independent test sets (training on 761 and testing on another 100 documents), we achieved the following: sensitivity=80%, specificity=88%, with a total proportion correct of 86%. However, this model performed poorly in general, that is, when evaluated via a cross-validation. More specifically, performance on some of the leave-out sets was poor with a sensitivity and specificity around 50% and thus not predictive at all.
We trained two neural networks (DL0 and DL1) and combined them to construct an aggregate predictor (DL1+0). We first report the topologies of the two component deep neural network models and then their independent performances, and finally, the performance of DL1+0.
DL0 was trained with 758 labelled documents: 657 documents that belonged to patients annotated as unsuitable and 101 that belonged to suitable patients. The input layer had 758 nodes (not related to the 758 documents; 758 is the number of input variables for DL0). The three hidden rectifier layers each have 200 nodes (we experimented with tanh layers and with several other topologies including a decreasing number of nodes, and more layers with no significant improvements), and the output layer used softmax so that there were two outputs, being a 0 or 1, that is, reject or accept. DL1’s input layer had 102 input nodes, but was trained with 100 0s and 101 1s (figures 2, 3).
In order to evaluate our models, we used a fivefold cross-validation (performance was stable over other cross-validations ranging from 5-fold to 20-fold), and we performed an independent data test set evaluation. Cross-validation is a standard practice, which theoretically determines how generalisable our models are—a protocol is used which leaves out a data set for testing, trains on the complement of the data and repeats this a number of times to generate statistics regarding sensitivity and specificity. The performance of each of these models is given in table 2 and table 3.
One can compute the specificity and sensitivity from the tables above. For DL0, the specificity is 97% and the sensitivity is 44.5%. In contrast for DL1, the specificity is 53% but the sensitivity is 89%. This means that we have one model that consistently performs well when classifying 0s and another model that performs well when classifying 1s. By experimenting with the topology of the neural network, it was possible to trade in a loss of specificity for DL0 to gain some sensitivity. It is worth mentioning that a model similar to DL0, which we shall call DL0_2, was trained that performed quite well in general. It had a specificity of 87% and a sensitivity of 75%. It was trained and tested on the same data as DL0 via a similar cross-validation process (refer Table 4).
As described, we used cross-validation to produce and tune a set of models that we combined into a set of DL1+0 models. For replication, we tested these models on a second, completely separate, independent test set of 103 documents (which included 25 true candidates, ie, documents labelled as 1) that were not included with the original 758 documents mentioned above. The DL1+0 algorithm yields as output a set of documents that correspond to patients that it considers suitable participants; next we report on how it performed on the second independent test set of 103 documents.
The sets of patients that DL1+0 identified as suitable participants ranged depending on the training of the neural networks. Many of the models generated returned predictions of 15 documents of which 13 were correct, giving a positive predictive value (precision) of 87%. Another set of models returned predictions for 22 out of the 25 possible suitable participants, 77% (ie, 17/22) of which were correctly identified as suitable, in terms of precision (positive predictive value). In practice, one could choose a model that would reveal many suitable candidates accurately, but that would miss many possible patients. Alternatively, one could use a model that returned more suitable candidates but it would include some patients that would not be suitable Table 5.
An actual output example is given here for two of the these models:
Input: 103 documents, 25 of which are annotated as suitable participants.
Output of DL1+0 (called suitable by DL1+0) = (41,43,44,45,46,47,48,55,56,57,58,60,62,66,67,70,72,73,74,75,77,99). Of these 22 documents, 62,66,67,75 and 77 were not annotated as suitable, which means that the output returned 17/22 (77%) correct calls.
Output of DL1+0 (called suitable by DL1+0_short) = (41,44,45,47,48,57,58,60,62,70,72,74,77,99,102). Of these 15 documents, 62 and 102 were not annotated as suitable, which means that 13/15 were correct calls.
DL1+0 is excellent at rejecting patients correctly with a worst-case score of 90% specificity, which occurs when sensitivity is 68%.
Though statistically this model appears very similar to the single-shot neural network model DL0_2, the user can be more certain of the reliability of the output list of recommended patients due to an increase in precision (positive predictive value). After several tests, DL1+0 consistently returns lists that are more conservative but more precise than DL1, DL0 or DL0_2 alone.
To summarise, we deidentified a corpus of EMR documents from a set of patients, annotated it using a set of inclusion and exclusion criteria, and used brute force and deep neural network approaches to phenotype potential research participants. Performance of the brute force method was inconsistent. We constructed a recommendation system by first training two deep neural networks, one that accurately recognises patients who are not suitable and another that accurately recognises patients who are suitable. We combined the two deep neural network models into a single model to augment a researcher’s ability to recruit suitable participants. By missing many potential participants, we have found that this algorithm can return document lists that are up to 87% accurate. This was validated on an independent test set after tuning each component with a fivefold cross-validation protocol.
The current investigation has several limitations. The most important potential limitation is the phenotype itself: in the DSM-5 field trials, the kappa for Major Depressive Disorder was 0.28 for both adult and child versions.27 Further, we were recently reminded that Major Depressive Disorder is an index of something and that we should not take an index of something as the thing itself.28 Psychiatric symptoms have successfully been extracted from EMR data on patients with serious mental illness,29 and this may be an alternative approach; to improve on a symptom-based phenotyping method, a network/complex dynamic system model may also be informative.30
We were not able to make use of structured diagnosis codes as are commonly available in most EMRs. It may be argued that discrete DSM or Systematized Nomenclature of Medicine (SNOMED) codes being available would render our deep neural network approach unnecessary; however, published evidence suggests that NLP/ML methods improve information extraction tasks in non-psychiatric phenotypes9 10 12 13 31 and depression.17 18 21
The poor performance of the brute force method led us to abandon this approach. The suspected reason for this fluctuation in performance is based on the understanding that the content of the EMR documents could vary substantially. Some documents have clear diagnoses, while others have clear narrative, and then others would be too ambiguous for the brute force approach to capture reliable information. Improving the PD and ND may help, but if these terms are too comprehensive, it will limit the types of patients that are recommended. An immediate step to improve this approach would be to apply some Bayesian methods—probabilistic methods that can capture a distribution of responses and make the procedure more flexible to variation.
Several supervised ML techniques are available and we used feedforward deep neural networks trained directly on document term matrices. It is important to note that we attempted to use singular value decomposition techniques to reduce noise but in our case it reduced the performance of our models. We should mention however that our algorithms already reduce noise by only considering words that occur above some threshold. Our modest sample size is the greatest contributor to the low sensitivity of our results. We used neural networks because we wished to experiment with a method that is truly generalisable for our task—our EMR documents are a heterogeneous and complex data set, representing several distinct psychiatric patient populations—a larger corpus may have yielded more accurate results. In total, there were 861 documents on 366 patients for the 6-month period. This limitation reminds us to treat this effort as a proof of concept. However, our models did replicate on a fully independent data set, suggesting our methods have some merit. In the future, we will use a larger training set in addition to a more powerful variant of neural networks known as recursive deep networks which have shown promise for natural language efforts.32 Future experiments will involve other ML techniques such as gradient boosting.33
The authors would like to acknowledge the support of Mr Chris Wakefield.
Funding University of Toronto McLaughlin Centre, grant number: MC 2014-18. This work was supported by a McLaughlin Accelerator Grant in Genomic Medicine (PW, JS).
Competing interests None declared
Provenance and peer review Not commissioned; externally peer reviewed.