Identifying phenotypic signatures of neuropsychiatric disorders from electronic medical records

Abstract

Objective

Mental illness is the leading cause of disability in the USA, but boundaries between different mental illnesses are notoriously difficult to define. Electronic medical records (EMRs) have recently emerged as a powerful new source of information for defining the phenotypic signatures of specific diseases. We investigated how EMR-based text mining and statistical analysis could elucidate the phenotypic boundaries of three important neuropsychiatric illnesses—autism, bipolar disorder, and schizophrenia.

Methods

We analyzed the medical records of over 7000 patients at two facilities using an automated text-processing pipeline to annotate the clinical notes with Unified Medical Language System codes and then searching for enriched codes, and associations among codes, that were representative of the three disorders. We used dimensionality-reduction techniques on individual patient records to understand individual-level phenotypic variation within each disorder, as well as the degree of overlap among disorders.

Results

We demonstrate that automated EMR mining can be used to extract relevant drugs and phenotypes associated with neuropsychiatric disorders and characteristic patterns of associations among them. Patient-level analyses suggest a clear separation between autism and the other disorders, while revealing significant overlap between schizophrenia and bipolar disorder. They also enable localization of individual patients within the phenotypic ‘landscape’ of each disorder.

Conclusions

Because EMRs reflect the realities of patient care rather than idealized conceptualizations of disease states, we argue that automated EMR mining can help define the boundaries between different mental illnesses, facilitate cohort building for clinical and genomic studies, and reveal how clear expert-defined disease boundaries are in practice.

Background

Fluid boundaries of mental illnesses

Over the course of the 20th century, scientists have struggled to refine the mental health classification system. Even as diagnostic manuals swell with new disorders and additional definitions of existing conditions, the underlying complexity of the mental illness landscape has continued to handicap effective diagnosis and treatment. Mental disorders are often characterized as distinct illnesses despite high rates of comorbidity, which indicate more fluid boundaries.¹ Some have argued that even the extensively updated DSM-V (fifth edition of the Diagnostic and Statistical Manual of Mental Disorders) does little to address the need of an objective approach to disease classification.^{2 3} Diagnostic criteria continue to evolve; new categories are added, and some traditional categories have recently been dropped.⁴

Because electronic medical records (EMRs) reflect the realities of patient care rather than idealized conceptualizations of diseases, they could potentially serve as a new source of information for this classification task. By revealing the phenotypic signatures of different disorders as they are observed and recorded by clinicians, EMRs could allow a more data-driven approach to the classification of mental illnesses.^5–7

Among mental illnesses, three related disorders—autism, bipolar disorder, and schizophrenia—have a particularly high impact on affected individuals and their families, and present a heavy economic burden for the healthcare system.^8–11 These disorders are complex, with a significant degree of shared clinical presentation, and may share common genetic origins.¹² For example, recent work exploring the genetic link between schizophrenia and autism has suggested that certain rare structural genetic variants, including copy number variants, play a role in both disorders.^{13 14} Bipolar disorder and schizophrenia have also been shown to overlap because of their shared presentation of certain ‘neurological soft signs’,¹⁵ as well as cognitive and memory impairments.^{16 17} Certain copy number variants are also common to the two disorders.¹⁸ Finally, case studies have suggested a genetic link between bipolar disorder and autism.^{19 20}

Our approach

In an effort to elucidate the shared and distinct phenotypic features of autism, bipolar disorder, and schizophrenia as they are discussed in clinical records, we mined the records of over 7000 psychiatric patients from Stanford Hospital and the Palo Alto Medical Foundation (PAMF). After annotating these records with concepts from the Unified Medical Language System (UMLS) and other ontologies, we searched for concepts that were enriched in each disorder relative to the baseline population at each facility. We also identified highly associated concept pairs in each disorder, and we applied dimensionality-reduction techniques to the individual patient records to determine variations in the clinical presentations for these disorders. We present our findings with a focus on the specific areas of overlap among the different disorders. We demonstrate that EMRs present a rich, objective and expert-sourced repository of information for studies seeking to delineate the clinical, phenotypic boundaries between important neuropsychiatric disorders.

Methods

Data sources

Our data consisted of EMRs from Stanford University's STRIDE (Stanford Translational Research Integrated Database Environment) database and PAMF. Stanford Hospital is a 613-bed teaching hospital and tertiary care facility; it admits over 24 000 patients per year and receives over 558 000 outpatient visits. In contrast, PAMF consists of community-based clinics throughout the San Francisco Bay area, covering roughly 700 000 patients across Alameda, San Mateo, Santa Clara, and Santa Cruz counties. Because Stanford Hospital is a referral facility, its psychiatric patient population might not be representative of the general population of patients with autism, bipolar disorder, and schizophrenia. We therefore considered data from both facilities in all parts of our analysis.

All of the data in this study were deidentified using the method described in US Patent Application 13 420 402. We used this technology to process clinical notes that contained protected health information, keep only the medically relevant concepts, and transform the data into a deidentified patient–feature matrix.^{21 22} The Stanford Institutional Review Board determined that the analysis of patient data deidentified in this way was not human subjects research, and the PAMF Institutional Review Board approved the study under an expedited review protocol for deidentified data. The technology used to deidentify the data is freely available to academics, and can be licensed non-exclusively to commercial entities.

Patient cohorts

Our initial patient cohorts consisted of patients who had received International Classification of Diseases (ICD-9) diagnosis codes of autism, bipolar disorder, or schizophrenia at any visit. These patients and their notes constitute the ‘Before pruning’ column in table 2. We then narrowed the cohorts, keeping patients who had received two separate diagnoses for the same disorder within 1 year and had more than 1 year between their earliest and latest notes. While there is no standard way to identify patients with a specific mental diagnosis based on EMR data alone, the best-performing algorithm for extracting depressed patients from EMRs requires two separate diagnoses and at least 1 year of follow-up.^{23 24}

Table 2.

Patients and records (notes) represented in the STRIDE and PAMF corpora, before and after pruning

	Before pruning		After pruning
	Patients	Notes	Patients	Notes
STRIDE, autism	2037	30 718	533	6598
PAMF, autism	1474	108 769	610	15 941
STRIDE, bipolar	5901	139 455	2946	55 700
PAMF, bipolar	5299	677 980	2296	129 980
STRIDE, schizophrenia	2198	45 292	270	1886
PAMF, schizophrenia	1018	136 758	449	27 937
STRIDE, total	10 136	215 465	3749	64 184
PAMF, total	7791	923 507	3355	173 858

PAMF, Palo Alto Medical Foundation; STRIDE, Stanford Translational Research Integrated Database Environment.

After identifying candidate patients, we extracted only those notes that were associated with patient visits. For STRIDE, the final set of notes included clinic notes, consultation notes, progress/discharge/transfer summaries, letters (such as referral letters) written by physicians, and patient histories/physicals. Pathology results and radiology transcripts were excluded. Detailed initial psychological evaluations were also excluded in the interest of patient privacy. For PAMF, the final set of notes included letters, history/physical notes, procedure notes, and problem/visit notes. Patient instructions and transcripts of phone calls were excluded. The pruned set of patients and notes constitutes the ‘After pruning’ column in table 2.

Annotation pipeline

The final set of notes was processed with the annotation pipeline described previously.^25–27 We used an optimized version of the NCBO (National Center for Biomedical Ontology) Annotator²⁸ to annotate the patient notes with concepts from 22 clinically relevant ontologies. The pipeline involved:

1. Generating a corpus of ∼5.6 million strings from 22 relevant ontologies and all trigger terms from NegEx²⁹ and ConText.³⁰

2. Pruning by term frequency and syntactic type information (eg, predominant noun phrases) from Medline to create a clean lexicon,^31–33 which was used for dictionary-based concept recognition to produce annotations.

3. Using NegEx and ConText rules to filter out negated terms and terms found in the family history.

4. Normalizing terms into concepts based on interontology mappings and semantic grouping by drug, disease, device, or procedure.

5. Normalizing drugs into ingredients using RxNorm,³⁴ so that (for example) ‘fluoxetine’ and ‘Prozac’ were both normalized to ‘fluoxetine’. Brand name mentions were normalized to their generic names, and combination drugs to their constituent substances.

Concept enrichment calculations

We first identified those concepts that were (a) significantly enriched among patients with a particular mental disorder relative to the general population and (b) not indicative of the patient population at a particular facility, but rather, indicative of the disorder itself. Because our six patient cohorts (table 2) were of different sizes, statistical power calculations were critical: we wanted to ensure that the minimum effect size required for concept enrichment was the same across all three disorders, as well as across both facilities. For each disorder, we carried out the following:

1. Randomly selected 1000 notes from patients with that disorder. This process involved first selecting a patient at random, then selecting one of that patient's notes (again at random). Thus patients with longer histories were not over-represented.

2. Randomly selected 1000 notes from patients without a mental disorder diagnosis at the same facility in the same fashion.

3. For each UMLS concept found in a note from a patient with the disorder, performed Fisher's exact test to determine if the term was significantly enriched among patients with the disorder (p<0.05).

4. Repeated steps 1–3 twenty times for each facility.

Using the statistical software package G*Power,³⁵ we determined that a cohort size of 1000 ‘cases’ and 1000 ‘controls’ would yield a power of 0.80 for detecting effect sizes of ∼6% difference in proportions between the cases and controls. We therefore considered a term ‘enriched’ for a particular mental disorder if Fisher's test found it so with p<0.05 in at least 80% of our trials from both facilities. It is important to note that this sampling procedure did not guarantee that each note in the sample came from a unique patient, even when the number of patients was much greater than 1000; it simply ensured that sampling was uniform across patients, regardless of length of history.

Finding associated concepts

We defined the clinical presentation of a disease along two dimensions: the concepts enriched for that disease, and the significant associations among those concepts. Power and effect size calculations were less straightforward for the association analysis because of the widely varying numbers of notes in which each concept occurred. We wanted to ensure that we did not see an increased number of significant concept–concept associations for a given disorder, such as bipolar disorder, simply because the larger number of notes for that disorder conferred greater power (the ability to detect associations with smaller effect sizes).

If we consider two concepts, C1 and C2, the ‘effect size’ was the difference in proportions of occurrence of C2 between notes where (a) C1 occurred and (b) C1 did not occur. If C1 and C2 were highly associated, we would expect C2 to occur in a high proportion of notes that contained C1 and a lower proportion of notes that did not contain C1. Initial power calculations revealed that, with an overall sample size of 1866 (the size of our smallest set of notes, for schizophrenia/STRIDE), we could observe an effect size of 8% with a power of 0.8 at a significance level of 0.05, even if the distribution of notes containing C1 versus notes without C1 were skewed as much as 4:1.

We used Fisher's exact test to find enriched associations between all pairs of concepts for each disorder where at least one of the concepts was among our set of 123 ‘enriched’ drugs or phenotypes for autism, bipolar disorder, and schizophrenia (figure 1). We required that the effect size be at least 8% for an association to be significant. This meant that, even if a given association was significant with p<0.05, it would not be accepted if the post hoc estimated effect size was not large enough to assure us that we would see the same association in the schizophrenia/STRIDE group with 80% power. Doing so reduced the chance that we would call an association significant for one disorder but not for a second disorder simply because the second disorder had a smaller sample size.

Figure 1.

Significantly enriched psychoactive drug terms for each disease, and their overlap.

Finally, for each pair of concepts, we actually measured two separate associations: the chance of observing C2, depending on whether C1 was present in the note, and the chance of observing C1, depending on whether C2 was present. The two effect sizes may be different, and the two directional associations were treated as separate in our analysis and viewed in our subsequent network visualization as directed edges between C1 and C2.

Network visualization and analysis were performed using Cytoscape 2.8.³⁶ A ‘neighbor’ of a node in a network is any node that is connected to it via an edge, ignoring edge direction. The ‘mean degree’ is the average number of neighbors of a node in the network. The ‘clustering coefficient’ is defined as the number of edges among neighbors of a specific node, divided by the total number of edges that could possibly exist among those neighbors, averaged over all nodes in the network. It is a measure of transitivity: how likely it is that ‘a friend of my friend is also my friend’.

Practically speaking, these network metrics provide different, though related, perspectives on similar phenomena. A high mean degree indicates the presence of a large number of significant associations among concepts for a particular disorder; the concepts are not distributed independently throughout patient notes, but tend to occur preferentially with certain other concepts: the higher the mean degree, the greater the number of significant associations. The clustering coefficient is related to the presence of community structure in the network.³⁷ Networks with high clustering coefficients are often distinguished by the presence of ‘hubs’: concepts with a disproportionately large number of connections.

Patient-level analysis

To visualize the overlap among patients with the three disorders, we consolidated each patient's record into a feature vector of length 123: one element for each of the 45 enriched drugs and 78 enriched phenotypes shown in figure 1 and figure 2. The value at each location was the number of mentions of the term in each patient's record, divided by the total number of notes in the patient's record. We normalized by the number of notes in the patient's record to ensure that two patients with similar distributions of concepts looked similar, even if one had a greater number of total notes (perhaps because of a longer follow-up time). We then performed principal components analysis³⁸ on the patient–concept matrix and graphed the patients’ locations along the first three principal components.

Figure 2.

Significantly enriched phenotypes for each disease, and their overlap.

Results

Patient demographics and dataset characteristics

Age and gender information for the population in our study is shown in table 1. Autism shows a bias toward younger ages, while there is substantial overlap of age ranges for schizophrenia and bipolar disorder. The gender ratios for the three disorders differ, with autism enriched for males and bipolar for females, while schizophrenia is more evenly distributed.

Table 1.

Demographic information about the study population in the two centers

	Sex ratio (male:female)	Age (median; range)
STRIDE, autism	80:20	15; 4–63
PAMF, autism	82:18	10; 2–53
STRIDE, bipolar	33:67	51; 14–102
PAMF, bipolar	34:66	49; 10–103
STRIDE, schizophrenia	51:49	56; 19–98
PAMF, schizophrenia	44:56	53; 11–96

PAMF, Palo Alto Medical Foundation; STRIDE, Stanford Translational Research Integrated Database Environment.

Table 2 shows the numbers of patients and notes in each dataset, before and after preprocessing. Schizophrenia was the least common disorder, and the schizophrenia/STRIDE cohort our smallest dataset, containing 270 patients and 1886 notes. In contrast, bipolar disorder was the largest cohort; the bipolar/PAMF dataset contained 2296 patients and 129 980 notes.

Enriched concepts

We grouped the significantly enriched terms for each disease into drugs and phenotypes (symptoms or clinical findings). The enriched set of drugs for each disorder is shown in figure 1, and the enriched phenotypes are shown in figure 2.

While the overall pattern of drug use in figure 1 is not surprising, it serves as a useful validation for the phenotype enrichment and concept association calculations. Note that because of our note sampling procedure, the number of drugs found enriched for bipolar disorder is not an artifact of its larger sample size: there are in fact 40 different drugs significantly associated with bipolar disorder, compared with 13 for schizophrenia and 13 for autism. Bipolar patients overlap with schizophrenic patients mainly because of their use of antipsychotics, such as haloperidol (Haldol), and they overlap with autistic patients mainly because of their use of antidepressants, such as sertraline (Zoloft), and stimulants, such as methylphenidate (Ritalin). There are no drugs associated with both schizophrenia and autism that are not also associated with bipolar disorder. If we interpreted these disorders solely in terms of their associated drugs, we might conclude that bipolar disorder is the most variable in its presentation, and that autism and schizophrenia each overlap with distinct aspects of bipolar disorder.

Figure 2 shows the overlap among the phenotypes associated with each disorder. There is substantial phenotypic overlap between bipolar disorder and schizophrenia: these patients are more likely to abuse alcohol and suffer from chronic health problems such as obesity, chronic obstructive pulmonary disease, neuropathy, and hepatitis. (For example, approximately 10–15% of patients with bipolar disorder are also chronically infected with hepatitis C.³⁹) These associations may reflect these patients’ more advanced age as well as various social and environmental factors associated with mental illness; elevated rates of nicotine dependence, for example, are well known in both schizophrenia and bipolar disorder.^{40 41} Epilepsy and mental retardation are both common to autism and schizophrenia but not bipolar disorder.

From a diagnostic standpoint, the most interesting findings are those phenotypic concepts uniquely enriched in each disorder. If we assume that some of these specific phenotypes lead physicians to prefer one diagnosis over another, we may consider them ‘signature’ concepts for each disorder. For example, enriched phenotypes unique to autism include developmental delays (eg, in speech) or certain repetitive behaviors such as tics. (Although the concept ‘corn of toe’ seems incongruous here, the terms ‘corn’ and ‘callous’ both map to it and these conditions could be the result of repetitive behaviors common in autistic children.) Enriched phenotypes in bipolar disorder include symptoms associated with depression and anxiety, such as migraines, irritable bowel syndrome, sleep disorders, and ulcers, as well as the core diagnostic criterion of bipolar disorder itself: mania. Finally, schizophrenic patients exhibit both paranoia and schizoaffective traits, but also suffer from chronic problems such as heart failure, cancers, dementia, diabetes, and hypertension. These probably reflect the advanced age of schizophrenic patients as well as the physical toll of living with schizophrenia.

Patient-level analysis

Figure 3 shows the result of the patient-level principal components analysis. Each dot represents a single patient. We see that there is significant overlap between schizophrenic and bipolar patients; in fact, the schizophrenic patient cluster is nearly encompassed by the bipolar cluster in all subplots. In contrast, autistic patients form a distinct cluster.

Figure 3.

Patient-level plots of the first three principal components (PC1 vs PC2 in (C), PC2 vs PC3 in (A), and PC1 vs PC3 in (B)). Autistic patients are shown in orange, schizophrenic patients in light green, and bipolar patients in dark turquoise.

The first principal component (PC1) is a vector in the space of covariates (drugs, phenotypes) from the original dataset; it captures the direction along which most of the variability in the dataset lies.³⁸ The second principal component (PC2) captures the direction of next highest variability (orthogonal to the first); the others follow in turn. It is therefore interesting to look at the first few principal components from our data to see which drug and phenotype terms contribute the most to each.

The major contributors to PC1 are, in order, ‘depressive disorder’, ‘anxiety disorders’, ‘mental disorders’, ‘drug abuse’, ‘sleeplessness’, ‘mood disorders’, and ‘bipolar disorder’. These contributing terms all have a positive sign, which means that the more each term occurs the more it pulls the associated patient along the positive PC1 axis. PC1 includes terms indicating the presence or absence of broad sets of conditions and could perhaps represent the variability in a patient's primary presentation. Bipolar patients extend beyond the rest along the positive PC1 axis; their records appear to discuss a wider variety of different potential diagnoses. (Notably, the term ‘bipolar disorder’ itself also plays a prominent role in this component.)

PC2 includes the following major contributors in the positive direction: ‘autistic disorder’, ‘developmental delay’, ‘pervasive development disorder’, ‘developmental disabilities’, ‘valproate’, and ‘valproic acid’, followed closely by ‘dextroamphetamine’ and ‘methylphenidate’. All of these terms are highly associated with autism. Conversely, the major negative contributors to PC2 are ‘heart failure’, ‘congestive heart failure’, ‘chronic obstructive airway disease’, ‘hypertensive disease’, ‘schizophrenia’, ‘anemia’, and ‘degenerative polyarthritis’. PC2 might be considered the ‘autism/schizophrenia axis’; patients are situated further along the positive axis if they exhibit autistic traits and further along the negative axis if they exhibit schizophrenic traits or chronic health disorders commonly associated with schizophrenia. PC2 explains roughly half as much variance in the original data as PC1; a patient can be autistic-like or schizophrenic-like, but not both. (Of course, much of this variability can be explained by the relative ages of the two groups of patients.)

The third principal component (PC3) contains the following major positive contributing terms: ‘major depressive disorder’, ‘depressive disorder’, ‘sleep apnea syndromes’, ‘duloxetine’, ‘bupropion’, ‘sleep apnea, obstructive’, and ‘anxiety disorders’. Major negative contributing terms include ‘valproic acid’, ‘valproate’, ‘bipolar disorder’, ‘haloperidol’, ‘mania’, ‘risperidone’, ‘schizoaffective disorder’, and ‘psychotic disorders’. We might conclude that PC3 represents a tradeoff between depressive/anxious symptoms and psychotic/manic symptoms. A patient sits further along the positive PC3 axis if he or she is depressed, has trouble sleeping, is anxious, or is taking duloxetine or bupropion (both antidepressants). A patient moves toward the negative axis if he or she is taking valproate/valproic acid (a mood stabilizer and anticonvulsant), an antipsychotic such as haloperidol or risperidone, or has been diagnosed with mania or psychotic symptoms. There is considerable overlap between autistic and schizophrenic patients along PC3, while bipolar patients exhibit high variance in this aspect of the clinical presentation.

After considering the terms that contribute the most to each principal component, therefore, the picture in figure 3 becomes clearer. First, we see that diversity of clinical presentation is greater for bipolar patients than autistic or schizophrenic patients: bipolar patients show a wide spread across the PC1 axis, and extend off in the positive direction. (Interestingly, inter-rater reliability studies of diagnoses from the initial DSM-V field trials showed a κ statistic of 0.52 for bipolar I disorder in children compared with 0.69 for autism spectrum disorder in children,³ indicating higher inter-rater agreement for autism vs bipolar diagnoses.) Second, we see that patients can also be organized along an ‘autism/schizophrenia axis’, PC2, and along a ‘depressive/psychotic scale’, PC3. Autism and schizophrenia exhibit considerable separation along PC2 but not PC3. Bipolar overlaps with schizophrenia, but not autism, along PC2, and extends much further than either schizophrenia or autism into ‘depressive territory’ along PC3.

Enriched associations

If we consider the significant associations for each disorder as a network of directed edges (arcs) between concepts, the autism network encompasses 204 unique concepts and 363 arcs. The bipolar network includes 1160 unique concepts and 4306 arcs. The schizophrenia network includes 219 unique concepts and 485 arcs. There were 25 concept–concept associations that occurred in all three disorders, two that occurred for autism and schizophrenia but not bipolar, 136 that occurred for bipolar disorder and schizophrenia but not autism, and 112 that occurred for autism and bipolar disorder but not schizophrenia.

The full network files are available as online supplementary material for this paper, and can be viewed using Cytoscape or similar software.

We visualized each disease in terms of its characteristic phenotypes and the associations among them ignoring drugs, as well as phenotypes that are merely side effects of drugs (figure 4). We removed drug names from the networks, as well as all phenotype concepts that were connected to the networks only through their interactions with drugs. Table 3 contains a summary of each network's properties.

Figure 4.

Network representations of phenotype–phenotype associations for (A) autism, (B) bipolar disorder, and (C) schizophrenia. The nodes represent phenotypic concepts, and the node sizes are proportional to the number of connections for each node (disregarding directionality). The edge widths are proportional to the strength of the interaction between the two concepts they connect.

Table 3.

A summary of some basic network parameters for the phenotype networks shown in figure 4

Parameter	Autism value	Bipolar value	Schizophrenia value
Clustering coefficient	0.154	0.306	0.249
Mean degree (ignoring edge direction)	2.699	5.177	4.074
Hubs (nodes with highest degrees)	Hypersensitivity Asthma Autistic disorder Epilepsy Eczema Otitis media Obesity Constipation Depressive disorder Sleep apnea syndromes	Depressive disorder Gastroesophageal reflux disease Bipolar disorder Anemia Hyperlipidemia Sleep apnea syndromes Hypothyroidism Sleeplessness Asthma Vomiting	hypertensive disease Schizophrenia Chronic obstructive airway disease Gastroesophageal reflux disease Asthma Anemia Heart failure Constipation Congestive heart failure Depressive disorder

For the purposes of this analysis, the networks are treated as undirected.

The differences among the three networks are apparent even by inspection. The network for bipolar disorder is the largest and has the highest clustering coefficient and the highest mean degree. It appears to form one connected ‘community’ of edges without any fragmentation. On the opposite end of the spectrum is autism, a much smaller network with the lowest clustering coefficient and lowest mean degree. The autism network consists of six separate fragments, and there exist nodes that, if removed, would fragment it further. From a topological standpoint, the autism network consists of phenotypic ‘communities’ that are loosely bound to each other, reflecting the heterogeneity of autism and the fact that it may actually be a group of disorders.⁴² Schizophrenia's network lies between these two extremes.

Discussion

Phenotypic signatures in cohort building

The annotation and statistical analysis methods described here serve two main purposes. First, recognition of ontological concepts in unstructured EMR text followed by statistical enrichment analysis identifies a set of UMLS concepts, and associations among concepts, that are indicative of one or more disorders of interest. This process defines phenotypic ‘signatures’ for disorders, enabling researchers interested in cohort building for clinical or genomic studies to identify patients that may meet the phenotypic criteria for a particular disease. Importantly, phenotypic signatures can be ‘built’ using EMR text from one facility and then applied to text from a different facility to identify phenotypically similar patients, even if specific diagnosis codes are applied differently at the two facilities. Such signatures could also play a crucial role in the development of a new ‘taxonomy’ of diseases, as suggested by the National Academy of Sciences.⁴³

Second, once the phenotypic signatures are identified, individual patients can be situated relative to the principal conceptual ‘axes’ within the overall landscape of a disorder. Considering an individual as a point in a high-dimensional feature space—where each feature is one of the ‘signature’ concepts for a disorder—analyses such as ours enable researchers to pick out ‘matched’ patients from different disorder groups, and to identify patients who are more or less representative of different disorders; both of which are useful steps in cohort building. Although we have focused on mental disorders, the methods described here can be applied to other disorder(s) of interest.

Related work

Several groups have explored ways to define phenotypic signatures for diseases from EMRs,⁴⁴ although most have used structured EMR data (diagnosis codes, etc) rather than unstructured text. For example, Hanauer et al⁴⁵ examined temporal associations among ICD-9 codes, and Kohane et al⁴⁶ showed that patients with autism have significantly increased rates of other comorbidities. Denny et al⁴⁷ were the first to conceptualize the PheWAS, or Phenome-Wide Association Study, where EMR data are used to identify specific phenotypes of interest, and then genetic data from those patients are examined to identify associated genetic variations. Data-driven phenotyping methods similar to ours are therefore essential to nuanced PheWAS studies for neuropsychiatric disorders. Finally, recent work on ‘EMR phenotyping’ has resulted in a web knowledge base that collects algorithms designed to extract patients with specific phenotypes from EMRs (http://phekb.org). Our approach is complementary to this effort; the annotation pipeline and enrichment statistics framework can be applied to develop phenotypic signatures for multiple diseases of interest.

Study limitations

As with any observational study, especially one where much about the underlying data collection process is unknown, ours suffers from several important limitations. Foremost among these is the difficulty we faced in validating our findings, since there is no such thing as a training set of all known UMLS concepts associated with schizophrenia, for example, or all known associations among different UMLS concepts for these different disorders. The most we could hope for was that a psychiatrist would deem our findings reasonable. We did present this work to several psychiatrists and other researchers interested in mental illness, but it would be impossible to ask an expert to produce a list of all relevant concepts and associations that exist so we could estimate our recall. Therefore, our study is necessarily somewhat qualitative and descriptive in nature.

In addition, it is possible that clinicians might base their descriptions of patients on what is in the DSM, in an attempt to provide support for a particular diagnosis. If this is the case, we might miss rarer associations with concepts that are not found in the DSM for a particular disorder. We may also have biased our results somewhat because of a lack of knowledge (as a result of privacy concerns) about the specialties of the particular physicians who wrote the clinical notes; an internist might record a longer and more detailed history for a patient who has seen a psychiatrist, for example, than for one who has not.

Future directions

Beyond cohort building and the identification of phenotypic signatures for specific diseases, our approaches could be extended to other problems related to EMR-based phenotyping. For example, the use of EMR-based text mining in combination with network-based analyses will likely have broad utility for uncovering new associations between clinical entities⁴⁸ and potentially in analyzing patient outcomes.⁴⁹ In addition, it is known that psychiatry patients have a higher mortality and morbidity from ‘usual’ causes, such as heart failure, diabetes, etc.⁵⁰ Knowing the strength of associations among comorbidities, as we have calculated in our association networks, would assist in devising care management protocols similar to those in Petri et al.⁵¹

Conclusion

By annotating, and then statistically exploring, the text of EMRs associated with mental illness, we have examined the phenotypic signatures of autism, bipolar disorder, and schizophrenia from three perspectives: the enriched concepts for each disorder, the networks of associations among those concepts, and the clustering of patients based on the concept mentions in their records. We have also examined the combinations of concepts that account for the most variance in patient records, uncovering a ‘diagnostic variability’ axis, an ‘autistic/schizophrenic’ axis, and a ‘depressive/psychotic’ axis. Our experiences analyzing these data underscore the importance of statistical power calculations to reduce the chances of introducing artifacts when comparing disease populations of very different sizes.

We have demonstrated that EMR mining can extract relevant drugs and phenotypes associated with three important psychiatric disorders, and can reveal reasonable phenotypic signatures. Examination of the first three principal components appears to isolate autism as a separate disorder, while revealing significant overlap between schizophrenia and bipolar disorder. We conclude that such automated EMR mining can elucidate the phenotypic boundaries between different mental illnesses, reveal how clear the defined diagnostic boundaries are in practice, assist in data-driven cohort building, and assist in the classification of psychiatric disorders in an objective manner.