Toward Reliable Symptom Coding in Electronic Health Records for Symptom Assessment and Research

Abstract

To date, symptom documentation has mostly relied on clinical notes in electronic health records or patient-reported outcomes using disease-specific symptom inventories. To pro vide a common and precise language for symptom recording, assessment, and research, a comprehensive list of symptom codes is needed. The International Classification of Diseases, Ninth Revision or its clinical modification (International Classification of Diseases, Ninth Revision, Clinical Modification) has a range of codes designated for symptoms, but it does not contain codes for all possible symptoms, and not all codes in that range are symptom related. This study aimed to identify and categorize the first list of International Classification of Diseases, Ninth Revision, Clinical Modification symptom codes for a general population and demonstrate their use to characterize symptoms of patients with type 2 diabetes mellitus in the Cerner database. A list of potential symptom codes was automatically extracted from the Unified Medical Language System Metathesaurus. Two clinical experts in symptom science and diabetes manually reviewed this list to identify and categorize codes as symptoms. A total of 1888 International Classification of Diseases, Ninth Revision, Clinical Modification symptom codes were identified and categorized into 65 categories. The symptom characterization using the newly obtained symptom codes and categories was found to be more reasonable than that using the previous symptom codes and categories on the same Cerner diabetes cohort.

Electronic health records (EHRs) have both positive and negative effects on healthcare, particularly medical diagnoses.¹ Nevertheless, EHRs provide a valuable source of data for medical research that complements traditional clinical trials, with large sample sizes and comprehensive clinical information.² EHR applications in clinical research include the discovery of disease patterns and comorbid conditions, cohort phenotyping, the identification of biomarkers and prediction of clinical outcomes, the quantification of the effects of interventions, and the detection of medication side effects and adverse events.³ EHRs contain structured data, such as laboratory results, diagnosis codes, and medications, as well as unstructured data such as clinical narrative text. Although a combination of multiple sources of data from EHRs can yield high performance for a task (eg, high accuracy for automatic phenotyping),^4,5 using standardized codes such as those of the International Classification of Diseases (ICD)⁶ or their regrouping by the Clinical Classifications Software (CCS)⁷ or according to phecodes⁸ allows for high throughput and supports large-scale studies given the breadth, portability, and ease of use of this data type.^9,10

Disease-specific symptom inventories developed to collect patient-reported outcomes (PROs) are commonly used for symptom assessment and research. Examples of widely used indices include the Memorial Symptom Assessment Scale (MSAS),¹¹ originally designed for cancer and also used for HIV, the HIV Symptom Index (HIV-SI),¹² and the Diabetes Symptom Self-care Inventory (DSSCI).¹³ One advantage of these indices is that they associate the contained symptoms with a scale for patients to report severity or distress related to each symptom. A major disadvantage is that use of these inventories requires questionnaire completion outside routine care, resulting in heavy burden on clinicians, patients, and researchers, and thus sample sizes are often limited. Moreover, each index contains only symptoms commonly related to one particular disease, limiting the scope for capturing data related to rare or uncommon symptoms.

With the availability of EHR data, symptom researchers have attempted to extract symptoms from clinical notes and unstructured data.^14,15 Although one can manually extract information from clinical notes for some individual patients, large-scale automatic information extraction from unstructured data for research purposes requires natural language processing (NLP). However, NLP may incur a high cost for model training and is not always feasible for large-scale studies, depending on levels of semantic processing, ranging from simple concept and named entity recognition to deep semantic analysis of text.^16–18 Additionally, symptoms extracted using NLP techniques, being either rule-based or machine learning–based, are not all correct and complete. That is because rule-based methods cannot correctly capture all possible symptom expressions, whereas the performance of machine learning methods is limited by the availability and correctness of training data, which require human annotations of clinical notes. Moreover, owing to data sensitivity and ethical issues with protected health information, deidentification of clinical notes is necessary and challenging.¹⁹ Importantly, most publicly shared EHR databases do not provide clinical notes, but only structured data such as diagnosis codes.

In addition, for symptom documentation and management using clinical notes, healthcare providers face challenges due to variations in symptom terminologies, which cause problems in reviewing, sharing, and aggregating symptoms entered by different persons and collected from different sources.²⁰ Therefore, like disease codes, symptom codes and their classification are desired for standardized and automatic symptom documentation, assessment, and research. However, so far, the ICD coding system for EHRs has been developed mainly for hospital and insurance billing purposes and thus focused on information such as disease diagnoses, medical procedures, drugs, and laboratory tests, but not symptoms.^21,22 Meanwhile, substantial human efforts to manually regroup ICD codes to capture more clinically meaningful concepts, such as in CCS or phecodes,^7,8 also do not focus on or better identify codes that represent symptoms.

The current ICD coding system used in most EHR systems in the United States is the clinical modification of ICD-10 (ie, ICD-10-CM).⁶ However, ICD-9 codes are still present in EHR databases created before ICD-10 was implemented and thus need to be considered when medical histories or retrospective studies are required. Many EHR databases available for research contain both ICD-9 and ICD-10 codes, where conversion of ICD-10 codes to ICD-9 codes is easy, not the other way around, because ICD-10 codes are more granular. In particular, the nationwide Cerner EHR database that we worked on contains healthcare records from 2000 to 2018, and thus, the majority of codes are in ICD-9-CM,²³ because ICD-10-CM went live only from October 1, 2015. For these reasons, this study started with ICD-9. ICD-9-CM was developed with multiple iterations, with an increasing number of codes, including parent codes and subdivided codes. Importantly, based on the most specific code use guideline, when an ICD code contains subdivided codes (ie, with additional digits), the subdivided codes should be used for diagnostic and billing purposes instead of the parent code.²⁴ The last ICD-9-CM system was version 32, effective from October 2014 and comprising 14 567 codes (to be used by this guideline) that have no subdivided codes.²⁵

Using the current ICD codes for EHR-based symptom assessment and research poses two main problems. First, although the code range 780–789 is designated for symptoms, these codes are not complete, and not all of them are correct for coding symptoms. In fact, the ICD-9-CM official guidelines for coding and reporting noted that the codes in that range contain many but not all codes for symptoms.²⁴ Meanwhile, some codes in the range do not represent symptoms, and some codes outside the range do represent symptoms.²⁶ Second, to our knowledge, except for one previous study,²⁶ there has been no effort to thoroughly determine which ICD codes are actually symptom related.

Importantly, symptoms can be reported only by a patient. As such, phenomena noted by someone other than a patient that indicate a change in normal functioning, sensation, or appearance due to disease are considered signs. If both a patient and someone else note a phenomenon, then it can be classified as both a sign and a symptom. Of note, some phenomena were previously considered symptoms (only noticeable by a patient), but with the introduction of technology can now be identified by another person and can thus be classified as signs, suggesting that the classification of signs and symptoms can change over time.²⁷

In a previous study,²⁶ symptoms associated with type 2 diabetes mellitus (T2DM) were characterized for the T2DM cohort in the Cerner database.²³ This database contains the healthcare records of around 69 million unique individuals across the United States with about 939 million ICD-9-CM and ICD-10-CM diagnosis codes. A multistep procedure was used to identify T2DM-related ICD-9-CM symptom codes from that cohort, after converting all the ICD-10-CM codes into the corresponding ICD-9-CM codes. The characteristics of T2DM-related symptoms, co-occurring symptom clusters, and their temporal patterns were derived based on the longitudinal EHR data of the cohort.

To date, diagnosis codes that represent symptoms in a general patient population have not been comprehensively identified and categorized. For large-scale EHR-based symptom assessment and research, especially when clinical notes are not available, identification and categorization of symptom diagnosis codes are essential. To our knowledge, and extending the previous study on the T2DM cohort in the Cerner database,²⁶ our study is the first attempt to identify and categorize ICD-9-CM symptom codes for a general patient population without specific disease restrictions.

METHODS

The Unified Medical Language System (UMLS) consists of software and data integrating many health and biomedical vocabularies and standards to enable interoprability between computer systems. Its three main components are a metathesaurus, a semantic network, and SPECIALIST lexicon and lexical NLP tools.²⁸ We employed the UMLS Metathesaurus, which is a large thesaurus of biomedical concepts from over 200 source vocabularies, including the ICD and other disease coding systems. Each concept is assigned the appropriate and most specific semantic type among the 135 semantic types of the UMLS semantic network. A total of 18 324 ICD-9-CM entry terms are included in the UMLS and are classified into 12 semantic types; of those, 15 279 are diagnosis codes, and the remaining terms are procedure codes.²⁹ Five UMLS semantic types were perceived to be most relevant to finding concepts representing signs or symptoms: Sign or Symptom; Mental or Behavioral Dysfunction; Finding; Injury or Poisoning; and Disease or Syndrome.³⁰ The UMLS definitions of these five semantic types and their corresponding ICD-9-CM codes are presented in Supplementary Table S1 (http://links.lww.com/CIN/A352).³¹

In this project, we aimed to (1) identify symptom codes from ICD-9-CM codes belonging to these five semantic types and group them into categories for a general patient population and (2) characterize symptoms in the T2DM cohort of the Cerner database as an application of the obtained symptom codes and categories. The previous study on only the Cerner T2DM cohort identified 931 ICD-9-CM symptom codes and 1087 3-digit codes and their subdivided codes that were deemed potentially non–symptom-related.²⁶ We used these results to highlight potential symptom codes and nonsymptom codes among the extracted codes from the UMLS to assist clinical experts to review and identify which codes are actually symptom related.

We identified symptom codes using four steps. First, the ICD-9-CM codes that belonged to the five aforementioned semantic types were automatically extracted from the UMLS. Next, the codes that were deemed symptom-related and non–symptom-related in the previous study²⁶ were automatically highlighted for clinical expert review. Then, two clinical experts with extensive experience in symptom science (both are PhD-prepared nurses and licensed advanced practice nurses with 20–30 years of clinical experience each; one expert is board certified in advanced diabetes management, and the other is a symptom science expert with over 40 symptom-focused peer-reviewed publications) carefully reviewed the descriptions of the codes to identify additional symptom codes, focusing particularly on the remaining unhighlighted codes. Finally, the two clinical experts grouped the identified ICD-9-CM symptom codes into categories based on their clinical meaning. Figure 1 presents a flowchart of this symptom code identification and categorization process.

Fig. 1.

The identification and categorization flowchart of ICD-9-CM symptom codes

To identify ICD-9-CM symptom codes and descriptions, we used the definition of symptoms as subjective phenomena reported by patients indicating a change in normal functioning, sensation, or appearance due to disease^32,33 and then applied the following rules:

1. If a symptom name is in the description of an ICD-9-CM code, we can confidently label this diagnosis code as a symptom (eg, fatigue); or

2. If no symptom name is in the description of an ICD-9-CM code, but the diagnosis cannot be made without a specific symptom being present, we can confidently label this diagnosis code as a symptom (eg, pityriasis rosea).

Once a diagnosis code was labeled as a symptom code, we then categorized those symptom codes into broad symptom categories based on their clinical meaning. We used the common symptom nomenclature based on ICD-9/ICD-10 code descriptions to name the symptom categories. The categories were based on traditional review of symptoms and aligned with the appropriate organ system (eg, skin, cardiovascular, gastrointestinal) or constitutional symptoms (eg, pain, fever). The “other” categories were used for symptoms that did not fit within a particular category but were closely aligned.

RESULTS

Extraction and Categorization of International Classification of Diseases, Ninth Revision, Clinical Modification Symptom Codes

A total of 12 987 ICD-9-CM codes belonging to the five aforementioned UMLS semantic types were extracted (Figure 1). Based on our proposed symptom identification rules, 1888 codes were asserted as symptom codes. The number of codes belonging to each semantic type after each step is presented (Figure 1). All identified symptom codes were then categorized into 65 broad symptom categories (Supplementary Table S2, http://links.lww.com/CIN/A352), with the specific ICD-9-CM codes and their number in each category. The broad symptom categories are listed in descending order based on the number of diagnosis codes contained in each category. We reidentified 130 of 931 previous symptom codes as non–symptom-related and 521 of 7488 previous non–symptom-related codes as symptom-related.

The largest symptom category was skin changes with 819 diagnosis codes, whereas the smallest ones were nosebleeds, heartburn, difficulty concentrating, vaginal dryness, dry mouth, nasal drainage, and wheezing, each of which contained only one diagnosis code. Fifty-two codes did not belong to the five targeted UMLS semantic types because they were from the previously identified 931 symptom codes.²⁶ Among those, six codes did not have an UMLS semantic type because they were not billable on/after October 01, 2015, and the other 46 codes belonged to four other UMLS semantic types: pathologic function (39 codes), anatomical abnormality (five codes), acquired abnormality (one code), and neoplastic process (one code). Eighty-five codes had descriptions that were miscellaneous or nonspecific. We classified those codes into other nonspecific symptoms. For example, 274.01 (“acute gouty arthropathy”) was classified into other musculoskeletal or neurologic symptoms, and 783 (“symptoms concerning nutrition, metabolism, and development”) was classified into other general symptoms. We excluded codes that could not be classified because of broad descriptions implying multiple symptoms, for example, 300.5 (“neurasthenia”) or 388.5 (“disorders of acoustic nerve”).

When checked against our identified symptom codes, in the 780–789 ICD-9-CM code range designated for symptoms, only 213 of the total 267 codes (ie, 79.77%) actually represented symptoms, whereas the other 54 ones did not, for example, 780.01 (“coma”) and 786.31 (“acute idiopathic pulmonary hemorrhage in infants”). In the larger 780–799 ICD-9-CM code range designated for symptoms, signs, and ill-defined conditions, only 226 of the total 443 codes (ie, 51.01%) were actually symptom-related, whereas the other 217 were not, for example, 790.01 (“precipitous drop in hematocrit”) and 793.11 (“solitary pulmonary nodule”). Meanwhile, 1662 of the 1888 ICD-9-CM codes identified as symptom-related in this project were not in the 780–799 range, eg, 032.84 (“diphtheritic cystitis”) and 719.44 (“pain in joint, hand”).

Comparison of the New Symptom Categories With Those in the Previous Study

Table 1 aligns the newly obtained symptom categories with those in the previous study,²⁶ which were only for the Cerner T2DM cohort. The new list is not limited to patients with T2DM and thus includes more symptom categories. The multiple categories in the previous list that corresponded to one category in the new list are separated by semicolons. There are 45 common categories between the two lists, six new categories each of which corresponds to more than one previous category (eg, change in sleep or difficulty sleeping corresponded to two categories, disturbed sleep and drowsiness/sleepiness), and 14 new categories not present in the previous category list (eg, numbness and tingling, joint stiffness).

Table 1.

Alignment of the new ICD-9-CM symptom categories to the previous symptom categories

New ICD-9-CM Symptom Categories for a General Population – 65 Items	Previous ICD-9-CM Symptom Categories for T2DM – 59 Items26
skin changes	skin changes
vision changes or eye problems	change in vision
pain	pain
headache	headache
change in sleep or difficulty sleeping	disturbed sleep; drowsiness/sleepiness
other musculoskeletal or neurologic symptoms	other musculoskeletal symptoms
bruising	bruising
change in hearing or tinnitus	hearing loss; tinnitus
swelling	swelling
depression or sadness	depression
anxiety	anxiety
change in urination, dysuria or urethral discharge	dysuria; urinary frequency; urinary retention; urethral discharge
nausea or vomiting	nausea/vomiting
difficulty remembering	trouble remembering
seizures	seizures
change in voice	change in voice
dizziness or syncope	dizziness
numbness and tingling
change in menstruation	change in menstruation
change in hair	change in hair
difficulty speaking	difficulty speaking
change in bowel patterns or diarrhea	change in bowel patterns
fever	fever
itching	itching
incontinence	incontinence
change in sexual interest or activity	change in sexual interest or activity
joint stiffness
change in sweating
shortness of breath	shortness of breath
trouble with learning
other general symptoms	other general symptoms; other abdominal symptoms
difficulty walking	difficulty walking
difficulty swallowing	difficulty swallowing
cough	cough
fatigue	fatigue
mouth sores
change in weight	change in weight
other gastrointestinal symptoms	other gastrointestinal symptoms; other GI symptoms
difficulty understanding or confusion	feeling confused
ear drainage
other head or neck symptoms	other head and neck symptoms
hemoptysis or change in sputum	hemoptysis
racing heartbeat	racing heartbeat
other mood symptoms
fingernail or toenail changes
other genitourinary symptoms	other urinary symptoms
trouble with coordination	trouble with coordination
change in appetite	change in appetite
pallor or flushing	flushing; pallor
breast changes
thirst	feeling thirsty
change in taste and smell	change in taste and smell
other cardiovascular symptoms	other cardiovascular symptoms
other mental health symptoms
chills or shaking	chills
other respiratory symptoms	other respiratory symptoms
weakness	weakness
sore throat	sore throat
nosebleeds	nosebleeds
heartburn	heartburn
difficulty concentrating
vaginal dryness
dry mouth
nasal drainage
wheezing	wheezing

Comparison of the New Symptom Categories With Commonly Used Symptom Indices

Various disease-specific symptom inventories have been developed for symptom assessment, management, and research using PROs. For cancer, at least 21 instruments have been identified as appropriate for clinical use,³⁴ including the 32-item MSAS.¹¹ Although originally developed for cancer symptoms, the MSAS has been used in many HIV symptom studies as well.^35,36 Specifically for HIV, commonly used instruments include the Sign and Symptom Check-List for Persons With HIV Disease,³⁷ the Istituto Superiore di Sanità-HIV Symptoms Scale,³⁸ and the 20-item HIV-SI.¹² The HIV-SI is considered the gold standard measure for clinical research related to HIV symptoms.^39–41 For T2DM, the DSSCI is a self-care inventory comprising 59 symptoms for psychometric testing with Mexican Americans.¹³ However, many diabetes-related symptoms are not included in that instrument, such as unsteady walking,⁴² joint stiffness,⁴³ and fingernail changes.⁴⁴

We compared our identified symptom categories for a general patient population with the three disease-specific symptom inventories used in cancer, HIV, and diabetes care and research—MSAS,¹¹ HIV-SI,¹² and DSSCI,¹³ respectively. Table 2 presents our identified symptom categories in comparison to those in the three inventories. For some symptom categories, multiple symptom items in the inventories corresponded to each of them, which are separated by semicolons. For example, two items in the MSAS, difficulty sleeping and nightmares, corresponded to the category change in sleep or difficulty sleeping in our ICD-9-CM symptom code classification. Meanwhile, the DSSCI contains susceptible to catching cold and infection, which we did not consider symptoms and thus did not identify their corresponding symptom categories. Moreover, unlike the three inventories, our symptom categories can be used for symptom assessment and research in a general patient population and are not limited to any particular disease. Of note, CCS and phecodes regroup ICD diagnosis and procedure codes into a manageable number of clinically meaningful categories, which effectively decreases the ICD code granularity.^7,8 However, unlike our categorization, they do not aim to identify which ICD codes are symptom-related, nor do they construct a comprehensive list of symptom categories for their own systems.

Table 2.

Alignment of the obtained ICD-9-CM symptom categories for a general patient population to three disease-specific symptom indices commonly used in cancer, HIV, and diabetes settings

ICD-9-CM Symptom Categories – 65 Items	Cancer Symptom Index (MSAS) – 32 Items11	HIV Symptom Index (HIVSI) – 20 Items12	Diabetes Symptom Index (DSSCI) – 39 Items13
skin changes		skin problems	dry skin; discolored skin; itchy skin
vision changes or eye problems			blurry vision; sensitive to light/noise
pain	pain	hand/foot pain; muscle/joint pain	burning sensation in feet
headache		headache	headache
change in sleep or difficulty sleeping	difficulty sleeping; nightmares	sleep trouble	problems in sleeping
other musculoskeletal or neurologic symptoms
bruising
change in hearing or tinnitus			sensitive to light/noise
swelling	swelling of arms or legs
depression or sadness	feeling sad	sadness
anxiety	feeling nervous; worrying	anxiety; body image	anxiousness; nervous; fidgety
change in urination, dysuria or urethral discharge	problems with urination; urinary accidents		urinating more than usual
nausea or vomiting	nausea; vomiting	nausea	indigestion or nausea
difficulty remembering		memory loss	memory loss
seizures
change in voice
dizziness or syncope	dizziness	dizziness	dizziness or light-headed
numbness and tingling	numbness/tingling in hands/feet		numbness or tingling of hands or feet
change in menstruation
change in hair	hair loss	hair loss	hair loss
difficulty speaking
change in bowel patterns or diarrhea	diarrhea; constipation	diarrhea	constipation
fever		fever
itching	itching		genital/vaginal itching
incontinence
change in sexual interest or activity	problems with sexual interest or activity	sex problems	loss of interest in sex; physical discomfort or problems performing sex
joint stiffness
change in sweating			sweating
shortness of breath	shortness of breath
trouble with learning
other general symptoms
difficulty walking
difficulty swallowing	difficulty swallowing
cough	cough	cough/SOB
fatigue	lack of energy; feeling drowsy	fatigue	tiredness
mouth sores	mouth sores
change in weight	weight loss	weight loss	weight loss; weight gain
other gastrointestinal symptoms	feeling bloated	bloating/gas
difficulty understanding or confusion
ear drainage
other head or neck symptoms
hemoptysis or change in sputum
racing heartbeat
other mood symptoms	feeling irritable		easily angry; irritability
fingernail or toenail changes
other genitourinary symptoms
trouble with coordination
change in appetite	lack of appetite	appetite loss	hungrier than usual; cravings
pallor or flushing			flushing
breast changes
thirst			intense thirstiness
change in taste and smell	change in taste
other cardiovascular symptoms
other mental health symptoms	I don’t like myself
chills or shaking			trembling
other respiratory symptoms
weakness			weakness
sore throat
nosebleeds
heartburn
difficulty concentrating	difficulty concentrating		trouble concentrating
vaginal dryness			vaginal dryness
dry mouth	dry mouth		dry mouth
nasal drainage
wheezing
			susceptible to catching cold
			infection

We identified 27 symptom categories that are not included in any of the three inventories. They comprise (1) 21 specific symptom categories (eg, bruising, seizures) and (2) six other non-specific or general symptom categories (eg, other musculoskeletal or neurologic symptoms, other gastrointestinal symptoms).

Comparison of the symptom categories identified in our list with those in the three inventories revealed several symptoms that may occur as a result of disease progression or as a side effect of treatment, but are missing in those inventories. Of note, the following symptoms can be associated with a diagnosis of diabetes: seizures (due to low blood sugar), difficulty speaking (due to low blood sugar), incontinence (due to urinary frequency associated with hyperglycemia), joint stiffness (due to inflammation), difficulty walking (due to neuropathy related to long-term diabetes), racing heartbeat (due to low blood sugar), fingernail or toenail changes (due to fungus caused by high blood sugar), and heartburn (due to gastroparesis).^45,46 Symptoms related to HIV include heartburn and other musculoskeletal or neurologic symptoms.³⁹ Symptoms related to cancers include bruising (due to low platelets/hematologic cancers), change in voice (related to head and neck cancers, mucositis/stomatitis), difficulty walking (due to chemotherapy-induced neuropathy), hemoptysis or change in sputum (due to lung cancer), breast changes (due to breast cancer), and sore throat (related to head and neck cancer or as a side effect of chemotherapy or radiation treatment).⁴⁷

Since the inception of disease-specific symptom inventories, symptom science has continued to evolve; thus, some of the symptoms that we identified are not present in the previously developed instruments. Additionally, our obtained symptom list is for a general patient population, rather than specific diseases, as with the previous instruments. Moreover, those symptom indices were designed for PROs to be manually collected using questionnaires; thus, they should not contain many items nor have fine category granularity. In contrast, by design, ICD codes are very detailed and specific to cover all possible conditions, including rare and uncommon symptoms. Therefore, our list includes more symptom categories, which could be further divided into subcategories when needed based on the ICD code hierarchical structure.

Characterization of Symptoms in the T2DM Cohort in a Large Nationwide EHR Database

We considered the T2DM cohort extracted by the previous study²⁶ from the Cerner database. First, we found that only 671 690 (about 0.4%) of the used 3- and 4-digit ICD-9-CM codes violated the most specific code use guideline introduced above²⁴ (Supplementary Table S3, http://links.lww.com/CIN/A352). That is, they were used while their subdivided codes were available. Therefore, it is safe to consider only the code list that follows this most specific code use guideline, which is shorter than the complete ICD-9-CM list.

Based on the identified symptom codes and categories, we characterized the symptoms in the Cerner T2DM cohort. In particular, we identified the 20 most prevalent symptom categories and compared the prevalence ranking results with those using the previous symptom codes and categories within this same cohort (Table 3). The prevalence ranking of a certain symptom category was based on the number of patients having symptoms in that category. In case multiple categories in the previous list correspond to one category in the new list, all the ranks of those multiple categories are shown.

Table 3.

The top 20 most prevalent new ICD-9-CM symptom categories in comparison with their ranks in the previous symptom categories on the same Cerner T2DM cohort

New ICD-9-CM Symptom Categories for a General Population – 65 Items	New Rank	Previous ICD-9-CM Symptom Categories for T2DM – 59 Items26	Previous Rank26
pain	1	pain	1
skin changes	2	skin changes	8
shortness of breath	3	shortness of breath	3
fatigue	4	fatigue	4
swelling	5	swelling	5
depression or sadness	6	depression	9
change in bowel patterns or diarrhea	7	change in bowel patterns	6
change in sleep or difficulty sleeping	8	disturbed sleep; drowsiness/sleepiness	7; 51
dizziness or syncope	9	dizziness	10
other musculoskeletal or neurologic symptoms	10	other musculoskeletal symptoms	19
nausea or vomiting	11	nausea/vomiting	12
change in urination, dysuria or urethral discharge	12	dysuria; urinary frequency; urinary retention; urethral discharge	21; 22; 23; 58
anxiety	13	anxiety	11
cough	14	cough	13
numbness and tingling	15
headache	16	headache	14
bruising	17	bruising	52
fever	18	fever	16
racing heartbeat	19	racing heartbeat	17
difficulty remembering	20	trouble remembering	35

On the one hand, the ranks of some symptom categories increased significantly compared with their ranks using the previous category list, namely, skin changes (rank increased from #8 to #2), other musculoskeletal or neurologic symptoms (#19 to #10), bruising (#52 to #17), and difficulty remembering (#35 to #20). This change can be attributed to our identification of more codes for these categories; thus, more patients had these symptoms. The rank change for these symptoms can also be attributed to frequent clinical manifestations among persons with diabetes. In fact, persons with T2DM often present with rashes, nonhealing wounds, or bruising as well as numbness, tingling, and sharp shooting pains in their feet as results of high blood glucose levels. These high glucose levels have been associated with “brain fog” or difficulty remembering.

On the other hand, a couple of symptom categories dropped out of the 20 most prevalent ones (Supplementary Table S4, http://links.lww.com/CIN/A352). In the previous study using the old category list,²⁶ heartburn ranked #2 and feeling confused (difficulty understanding or confusion in the new list) ranked #15 among the identified symptoms of T2DM patients. As we continued to work with the codes and further refined our definitions of symptoms, we found that several of the diagnosis codes represented diagnoses, not symptoms. Based on the aforementioned rules that we used to define symptoms, we determined several ICD-9-CM codes that we could not confidently identify as symptoms. Therefore, we removed codes 530.11, 530.81, and 536.8 from the heartburn category, dropping it from a rank of #2 to #49. Similarly, we removed codes 293, 293.0, 293.1, 298.2, 780.02, 780.1, and 780.97 from the feeling confused category, dropping it from #15 to #54.

Clinically, these changes appear reasonable, and our new rankings reflect a more accurate representation of the most prevalent symptoms among persons with T2DM. For instance, the rank change of heartburn reflects the fact that although heartburn can be a symptom, it can present with a wide variety of other symptoms, including burning pain in the chest, difficulty swallowing, passing gas, or nausea or vomiting. The nausea or vomiting category is ranked #11 in our new list, and passing gas, which falls within the change in bowel patterns or diarrhea category, is ranked #7 in the new list. For the feeling confused category, clinically, patients are more likely to report difficulty remembering, which is ranked #20 in our new list, whereas in the previous list, feeling confused contained codes for altered mental status, hallucinations, and delirium.

DISCUSSION

Our findings provide the first version of ICD-9-CM symptom codes that can be used as a starting benchmark for coding and extracting symptoms in EHRs to support symptom assessment and research. The identified symptom categories can also be used to analyze and characterize symptoms in a large general patient population or in cohorts of patients with specific diseases, which can then help identify risk factors and build prediction models for diseases.⁴⁸ For high-throughput automated phenotyping, previous studies combined ICD codes and lightweight NLP concepts extracted from clinical notes.^49,50 The newly identified symptom codes could be used for the development of symptom-enhanced phenotyping algorithms. Furthermore, as a byproduct, our study created a list of nonsymptom codes as negative examples beside a list of symptom codes as positive examples. The symptom codes and nonsymptom codes with their ICD descriptions can be used to train supervised machine learning models to help automatically identify symptom codes in other large diagnosis coding systems such as ICD-10-CM, for which manual review and identification would incur high labor costs.

Using the UMLS semantic types of the ICD-9-CM codes facilitated the identification of symptom codes in two ways. First, by considering only codes that belong to the five aforementioned UMLS semantic types, we could narrow down the set of diagnosis codes from over 15 000 codes to fewer than 5000 possible symptom-related codes to focus on. Second, within this focused set of codes, 1825 codes shared 890 UMLS concept unique identifiers (CUIs) with other codes. Supplementary Table S5 (http://links.lww.com/CIN/A352) presents the codes and the UMLS CUIs they share, for each of the five UMLS semantic types of discourse. For each subset of codes that have the same UMLS CUI, one needs to review only one code in the subset because the codes should consistently be either symptom-related or non–symptom-related. Thus, among those 1825 codes, we had to review only 890 representative codes corresponding to those different UMLS CUIs.

Study Limitations

First, owing to tremendous effort involved in manually reviewing codes, not every ICD-9-CM code and description could be examined. We focused instead on the codes that were potentially symptom-related for clinical expert review. Nevertheless, based on the five UMLS semantic types most relevant to signs or symptoms, our obtained list should cover a majority of ICD-9-CM symptom-related codes. Second, UMLS uses the most up-to-date ICD-9-CM version (the last ICD-9-CM update was in 2015); thus, the current UMLS version does not include a few codes in the older ICD-9-CM versions. Third, manual symptom identification and categorization could be susceptible to subjective biases. Fourth, owing to the coding practices of healthcare providers, the Cerner EHRs could have inaccurate or missing diagnosis codes that might affect our symptom characterization of the Cerner T2DM cohort. This limitation is true for any research using EHRs in general, not just for symptom codes but also for billable diagnosis codes. Last, although characterization of symptoms within the Cerner database demonstrated an application of the obtained ICD-9-CM symptom codes and categories, the results of symptom prevalence may not be generalizable to other patient populations.

CONCLUSIONS

So far, coding of symptoms in EHRs has not been given adequate attention compared with other clinical aspects such as diseases, medical procedures, drugs, and laboratory tests. Developing a comprehensive and reliable symptom coding system is important for symptom recording, reporting, and monitoring in EHRs, as well as for automatic symptom sharing, assessment, and research.

This work is an initiative to develop the first version of ICD-9-CM symptom codes and their categories for a general patient population. We chose to start with ICD-9-CM because it is important for longitudinal retrospective analysis of EHRs, when ICD-9-CM was used, and manual identification of symptom codes from a larger diagnosis coding system like ICD-10-CM would be labor intensive. The results include a list of nearly 1900 ICD-9-CM symptom codes grouped into 65 broad categories, which could be straightforwardly divided into subcategories based on the ICD code hierarchy. We have found new symptom codes representing many symptom categories that are missing from the originally designated ICD-9-CM symptom code range and commonly used disease-specific symptom indices.

The newly identified symptom codes and their categories were applied to characterize symptoms of the T2DM cohort in the nationwide Cerner EHR database. These symptom codes and categories could be used for symptom assessment and research using other EHR databases for cohorts representing any patient population as well. With a comprehensive list of symptom codes, the ability to automatically and rapidly find symptoms in patients allows for efficient symptom assessment and prediction of symptom-related health outcomes using large EHR databases. Toward this end, besides a rich set of symptom codes, health providers’ awareness of the usefulness of symptom coding needs to be addressed and improved, so that they will manually add symptom codes to each encounter as they have done with disease codes and other codes in EHRs.

For future work, first we suggest a Delphi study and the development of a tool with an application program interface for crowdsourcing to get feedback and consensus from subject experts to refine and update the obtained list of symptom codes and their categories. Second, we suggest evaluating improvements in symptom extraction using our newly identified symptom codes in comparison with the originally designated ICD-9-CM symptom code range, against ground-truth symptoms in an EHR database that contains clinical notes. Third, one can apply a combination of the identified symptom codes and other information encoded in laboratory test results, medications, clinical notes, and the like in EHRs to increase the accuracy of patient phenotyping and health outcome prediction. Finally, our identified symptom codes can be used as training data for machine learning to automatically identify symptom codes from larger diagnosis coding systems such as ICD-10-CM.