Natural Language Processing with Machine Learning Methods to Analyze Unstructured Patient-Reported Outcomes Derived from Electronic Health Records: A Systematic Review

Abstract

Objective:

Natural language processing (NLP) combined with machine learning (ML) techniques are increasingly used to process unstructured/free-text patient-reported outcome (PRO) data available in electronic health records (EHRs). This systematic review summarizes the literature reporting NLP/ML systems/toolkits for analyzing PROs in clinical narratives of EHRs and discusses the future directions for the application of this modality in clinical care.

Methods:

We searched PubMed, Scopus, and Web of Science for studies written in English between 1/1/2000 and 12/31/2020. Seventy-nine studies meeting the eligibility criteria were included. We abstracted and summarized information related to the study purpose, patient population, type/source/amount of unstructured PRO data, linguistic features, and NLP systems/toolkits for processing unstructured PROs in EHRs.

Results:

Most of the studies used NLP/ML techniques to extract PROs from clinical narratives (n=74) and mapped the extracted PROs into specific PRO domains for phenotyping or clustering purposes (n=26). Some studies used NLP/ML to process PROs for predicting disease progression or onset of adverse events (n=22) or developing/validating NLP/ML pipelines for analyzing unstructured PROs (n=19). Studies used different linguistic features, including lexical, syntactic, semantic, and contextual features, to process unstructured PROs. Among the 26 NLP systems/toolkits we identified, 14 used rule-based NLP, 8 used hybrid NLP, and 4 used ML algorithms embedded in NLP.

Conclusions:

This study supports the potential utility of different NLP/ML techniques in processing unstructured PROs available in EHRs for clinical care. Though using annotation rules for NLP/ML to analyze unstructured PROs is dominant, deploying novel neural ML-based methods is warranted.

INTRODUCTION

Patient-reported outcomes (PROs) provide information about a patient’s physical, psychological, somatic symptoms, daily functional status, health-related quality-of-life (HRQOL), and satisfaction with healthcare services that facilitate clinical decision-making and outcome evaluation [1]. PROs are conventionally assessed through validated questionnaires or semi-structured interviews. However, these methods may not capture the full patient experience as their responses are bound by predetermined items. Additionally, it is challenging to collect PRO data from patients during time-limited clinical encounters [2]. Therefore, finding effective approaches to assess, retrieve, and extract already available, unstructured PRO data from alternative sources, e.g., medical notes in electronic health records (EHRs), is important [3].

PRO data in EHRs are often stored in an unstructured and free-text format (e.g., symptom narratives in a physician note) and cannot be directly used in clinical tasks (e.g., disease prediction, classification). Natural language processing (NLP) and machine learning (ML) can turn unstructured PROs into a quantitative or structured format for clinical use [4]. The application of NLP/ML techniques in clinical settings includes text pre-processing (e.g., tokenization, lemmatization, or stemming from the corpus), linguistic feature extraction for clinical narratives (e.g., encoding, affirmed/negated expressions), and clinical applications [5]. Given the significant expansion of EHR ecosystems and the development of novel NLP/ML techniques over the past decade, the interest in processing unstructured PROs from clinical narratives through automatic or semi-automatic NLP/ML pipelines of free-text-based PROs is emerging [6, 7].

Previous reviews have described the application of NLP/ML methods for automatic extraction of non-PRO clinical narratives (e.g., disease progression, adverse drug reactions, medications, and treatments) from EHRs [5, 8], yet review studies of NLP/ML techniques on unstructured PROs are limited. One review article investigated text-mining NLP techniques to extract symptom information through patient-authored data collected from social media (e.g., Twitter, WebMD, and Reddit) [9]. Another review article collected the applications of NLP in analyzing symptom-only data documented in EHRs [10]. However, these studies mainly focus on the traditional rule-based NLP methods to process symptom data, followed by non-neural ML-based classifiers (e.g., support vector machine, logistic regression classifier), rather than novel neural network (e.g., CNN, RNN) or large language models (e.g., GPT, BERT) to analyze associations of PROs and clinical outcomes.

With the recent advances in NLP/ML techniques, this study aimed to summarize research applying NLP/ML for processing and analyzing unstructured PRO data collected in EHRs. Specifically, we evaluated studies that analyzed unstructured PROs for clinical care or research, the type of unstructured PRO data, and the uniqueness of NLP/ML systems/toolkits and techniques to process unstructured PROs. In contrast to previous review studies [9, 10], the findings from this study will improve our clinical insights of using NLP/ML techniques to process EHR-based unstructured PROs in a broader category (including symptom, functioning, and quality-of-life), together with other clinical parameters for clinical application.

METHODS

Data Retrieval

We searched studies written in English between January 1^st, 2000 and December 31^st, 2020 through PubMed, Scopus, and Web of Science. Following the guideline of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses, we identified 690 studies from PubMed, 739 from Scopus, and 886 from Web of Science (Figure 1). The search strategies are reported in Supplementary Table S1.

Figure 1.

Flow diagram of included articles

Article Selection

Among the 2,318 studies identified from these three databases, 979 non-duplicate studies were retained for the title and abstract screening. Studies were considered for inclusion if they 1) focused on unstructured PRO data in EHRs or medical notes, and 2) used NLP techniques or applications accompanied by ML algorithms to extract or process unstructured PRO data. Studies were excluded if they 1) did not apply NLP/ML techniques, 2) were non-empirical studies (e.g., case reports, commentary), 3) were non-EHR-based studies (e.g., patient-authored data collected from social media), 4) were survey-based studies containing quantitative PRO data, 5) were previous systematic review studies or E-pub ahead of print, and/or 6) focused on non-chronic disease topics (e.g., infectious disease/vaccination). Based on these criteria, the first author (JAS) and the senior author (ICH) independently reviewed the title and abstract of all 979 studies retrieved from the literature search and retained 151 studies. Subsequently, the same two authors reviewed the full‐text articles, resolved any discrepancies, and selected 79 studies for inclusion in this study (Figure 1).

Data Extraction and Summary

We implemented the following steps to collect data from the original articles: 1) the first author (JAS) manually extracted data from the 79 studies and documented information in a database, 2) the senior author (ICH) reviewed all of the extracted information in the database, 3) both co-authors met to confirm the extracted data through reviewing each of the original articles included in this study, and 4) the second author (XH) adjudicated the discrepancy raised by the first and senior authors, and reviewed information summarized in the tables for identifying any errors. For each selected study, the characteristics of the study sample, the objective of the study, the type/format, size, and unit of unstructured PRO data, specific PRO content, and NLP/ML systems/toolkits and linguistic features used to process PRO data were abstracted and reported.

RESULTS

Study Characteristics

Supplementary Table S2 displays the characteristics of 79 studies selected for inclusion in this systematic review. These studies included various participant sample sizes, ranging from 22 to 267,855, and different disease diagnoses, including any type (n=24) [7, 11–33], chronic disease (n=17) [6, 34–50], mental illness (n=23) [51–73], cancer (n=8) [74–81], and others (n=6). The majority of studies (n=73) focused on adults, six on pediatrics [19, 25, 38, 54, 82, 83], and one on both adult and pediatric patient populations [7]. Most of the PRO narrative data were primarily obtained from inpatient or outpatient EHRs (n=75), while several studies relied on the open data repositories/resources (e.g., MIMIC-III; n=4) [24, 33, 35, 50] and the national database (e.g., Taiwan’s National Health Insurance Research Database; n=1) [72]. Narrative data from both inpatient settings (e.g., admission notes, discharge summaries, nursing narratives, emergency department documents, intensive care unit reports; n=57) and outpatient settings (e.g., primary care documents, psychiatric evaluation notes; n=35) were generated by healthcare professionals. The number of PRO narrative documents/grammatical units for NLP/ML analysis ranged from 100 to 4.2 million. Approximately 85% of the studies were based on English-based EHRs (USA: 66%, UK: 15%, and Canada, Island, and Australia: 4%), 10% from Chinese-based EHRs (China: 7% and Taiwan: 3%), and 5% from EHRs with other languages (Danish: 3%, Swedish: 1%, and Egyptian Arabic: 1%).

The structure of the free-text PRO documents included keywords/phrases, sentences, paragraphs, or the entire document/medical note. Different vocabulary systems representing standardized clinical terminologies or nomenclatures (e.g., SNOMED-CT (n=10) [35, 42–44, 47, 56, 57, 60, 73, 78], UMLS (n=17) [17, 25, 26, 31, 33, 36, 42–45, 47, 59, 64, 74, 76, 82], LOINC (n=1) [18], DSM (n=3) [59, 60, 73], ICD-codes (n=26) [11, 15, 17–20, 23, 28, 34, 36, 38, 40, 42, 43, 50, 53, 54, 61–63, 67, 68, 83–85]) were used to extract the representation of unstructured PROs. Among 37 studies that used a rule-based NLP approach, 14 adopted an extant rule-based NLP system, and 21 created study-specific/custom rules to extract the representation of unstructured PROs. Additionally, 20 studies used clinical/medical dictionaries to incorporate the standard clinical terminologies (e.g., LOINC, UMLS, SNOMED-CT) or created custom or study-specific terms/concepts [12, 14, 16, 28–31, 37–40, 55, 60, 64, 69, 72] for extracting unstructured PROs.

Main Objectives of the NLP/ML Application in Analyzing PRO Data

Table 1 summarizes the features of studies that used NLP/ML techniques to process unstructured PRO data. Most of the studies used NLP/ML algorithms to identify or extract unstructured PROs from clinical narratives (n=74 studies), map or classify extracted PRO information into specific PRO domains (n=16), phenotype unstructured PROs to capture specific PRO features (n=9), and classify co-occurrence of multiple PRO problems (e.g., symptom clustering; n=5). Some studies used NLP/ML techniques to analyze unstructured PRO to predict the risk of disease progression or adverse medical events (n=22), develop/validate NLP/ML pipelines for analyzing unstructured PROs (n=19), test associations with clinical outcomes (n=5), and stratify or group patients for delivering tailored healthcare services per distinct patterns of PRO problems (n=3).

Table 1:

Purpose of NLP/ML applications for analyzing PRO data among the 79 studies included in the systematic review^*

Classifications	Specific roles of NLP/ML tasks	Task description	N	%
Information extraction/text identification			74	93.7
	PRO content detection, identification, extraction	Detect or identify PRO keywords or terminologies from free text	47	59.5
	PRO annotation	Perform semi-automated or manual annotation for PROs in free text	37	46.8
	PRO affirmation/ negation	Declare whether symptoms or symptom-related outcomes exist or equivalent expression or negative statement for having symptoms	38	48.1
	Vocabulary mapping	Map or assign PROs or PRO-related vocabulary words to appropriate indexes or labels	13	16.5
Classification/phenotyping/clustering			26	32.9
	PRO classification	Assign or classify extracted PROs into specific categories	16	20.3
	PRO phenotyping	Indicate specific characteristics of single or multiple PROs features	9	11.4
	PRO clustering	Identify two or more PROs that are related to each other or co-occur	5	6.3
Develop or validate NLP/ML pipelines			19	24.1
	Development of NLP/ML pipelines	Develop new NLP/ML pipelines or build NLP software	10	12.7
	Evaluation/validation	Evaluate and validate the performances of NLP system/pipeline	12	15.2
Risk prediction or stratification for clinical outcomes			25	31.6
	Risk prediction	Predict the risk of outcomes using extracted PROs based on unstructured narratives	22	27.8
	Risk stratification	Identify the right level of care and services for distinctive subgroups of patients.	3	3.8
Investigate associations between PROs and clinical outcomes			5	6.3
	Relationship detection	Detect semantic associations or relationships between unstructured PROs	5	6.3

^*Each study may include multiple study purposes and NLP/ML tasks.

Domains of Unstructured PROs Included in EHRs

Supplementary Table S3 displays the domains of unstructured PROs documented in EHRs that were extracted by NLP/ML approaches. The most popular documented PRO domains were psychological symptoms (e.g., anxiety, depression, stress; 57.0% of 79 studies), followed by general symptoms (e.g., fatigue, pain, insomnia; 48.8%), metabolic/hormonal symptoms (e.g., obese, weight loss; 28.8%), and physical symptoms in the digestive/gastrointestinal (e.g., bloating, constipation, diarrhea; 42.5%), nervous (e.g., arousal, attention problem, dysarthria; 41.3%), respiratory (e.g., cough, sneezing, dyspnea; 37.5%), cardiovascular (e.g., cardiac problem, angina, swelling of arms and legs; 28.8%), and dermatologic (e.g., itching, edema, rashes; 20.0%) organ systems. Several PRO domains (e.g., symptoms in the head/neck, musculoskeletal, urinary, sexual/reproductive organ systems; physical and social functioning) were documented by <20% of the studies.

Types and Methods for Linguistic Features

Table 2 provides examples of linguistic feature types derived from unstructured PROs: 1) lexical, 2) syntactic, 3) semantic, and 4) contextual features. Lexical features address the word-level characteristics of clinical narratives primarily using n-grams (e.g., uni- and bi-grams) representing the concept of unstructured PROs [51, 57, 61, 74]. Syntactic features address the phrase, clause, sentences, and part-of-speech tagging [12, 31, 45, 56, 82]. While syntactic features represent grammatical patterns (e.g., noun and adjective/adverb phrases) [31, 45], semantic features focus on the meaning of words and terms, typically defined in a custom dictionary, controlled medical terminology, or lexicon (e.g., LOINC, SNOMED-CT, UMLS) [31, 39, 51, 59, 62]. Representing the relevance of specific words to individual texts, as opposed to their prevalence in a corpus, is another method used to process semantic features and semantic keywords for clustering the symptoms [62]. Contextual features address the relative or absolute characteristics of unstructured PROs by considering the linguistic components (i.e., words, phrases, and sentences) neighboring around words or phrases of interest rather than searching keywords of unstructured PROs [6, 7, 22, 51, 76]. One study shows that subject terms (e.g., “mother”, “patient”), negation terms (e.g., “does not”), hypothetical terms (e.g., “if”), temporal terms (e.g., “previously”, “when”), and termination terms (e.g., “however”) were examples of contextual analytic features [39]. Another study labeled the contextual feature as “if-clause” for the text “I recommended nitroglycerin if he should develop chest pain”[22].

Table 2:

Examples of frequently used linguistic features

Levels of linguistic features	Description	Methods	Some example methods explained in selected studies
Lexical features	Numerical characteristics of tokens in text documents, such as token count and length.	N-gram (e.g., uni- and bi-grams), capitalization (uppercase, title case), stemming, lemmatization, stopwords removal, lexicon, word embeddings	“Stop-words removal (e.g., “is”, “an”, “the”, etc.), stemming, and number to string conversation.” [Banerjee, 2019] “Lexical variances in the extraction rules [i.e., misspellings (e.g., obese* instead of obsessive)].” [Chandran, 2019] “N-grams represent concepts of serious mental illness symptomatology.” [Jackson, 2018] “Text processing included lower casing; removal of punctuation, stop words, and numbers; word stemming; and tokenization.” [Obeid, 2020]
Syntactic features	Patterns of sentence structures defined by language grammar.	Part-of-speech (POS) tags, constituency grammar, dependency grammar	“POS tagger and multi-word term identification to identify symptoms and non-symptoms were used.” [Divita, 2017] “POS tags in conjunction with knowledge engineering features generated to build a sentence classifier.” [Jackson, 2017] “Syntactic phrases representative of patients’ functional status including noun phrases (e.g. “patient”), prepositional phrases (e.g. “with pain”), and adjective/adverb phrases (e.g. “very tired”) using two reference standards.” [Pakhomov, 2011] “Syntactic patterns of concept phrases were mined from continuous, non-permuted forms of synonyms, and these patterns were used to detect discontinuous and/or permuted concept phrases.” [Torii, 2018]
Semantic features	Linguistic units of meaning-holding components that represent word meaning, such as lexicon definitions, dependency between tokens, and semantic networks.	Semantic definitions from lexicons (LOINC, SNOMEDCT, UMLS, etc.), relative temporal words (next, later, until etc.), absolute temporal expressions (a.m., p.m., etc.), meaning of the numbers (doses, levels), deidentification, topics of the section	“Semantic variances in terms of obsessive and compulsive in the extraction (alternative meanings beyond their definition in the context of Obsessive Compulsive Symptoms (OCS).” [Chandran, 2019] “Semantic keywords identifying the Altered mental status cluster of symptoms in the context of pulmonary embolism.” [Obeid, 2019] “UMLS semantic networks which are relevant to clinical findings were used” [Torii, 2018] “The symptom dictionary was based on UMLS, which includes a semantic network.” [Le, 2018]
Contextual features	Linguistic neighboring components (e.g., word, phrase, or sentence) of tokens or sentences that represent similar semantic meanings.	Negation/affirmation, complex temporal relations, discourse structure, line position, order of sections, implicit context dependent information, feature representations from pre-trained neural embeddings	“Distinguishing between instances where a patient is described as experiencing a particular symptom from instances where the texts state that the patient is not experiencing that symptom, or where it is someone else (e.g. a friend or relative) who is experiencing that specific symptom.” [Chandran, 2019] “The “conditional” context label is considered when the term is mentioned in the following context (e.g., “I recommended nitroglycerin if he should develop chest pain”).” [Pakhomov, 2008] “Depending on the context, weight gain could indicate either fluid accumulation because of worsening heart failure or an improvement in appetite because of decreased gut edema associated with a higher dose of diuretics.” [Leiter, 2020] “Subject terms (e.g., ‘mother’, ‘patient’), negation terms (e.g., ‘does not’), hypothetical terms (e.g., ‘if’), temporal terms (e.g., ‘previously’) and termination terms (e.g., ‘however’).” [Iqbal, 2017]

NLP Systems/Toolkits for Processing Unstructured PROs

Table 3 displays the NLP systems/toolkits used to process unstructured PROs in EHRs, inclusive of 15 generic (e.g., ConText, FMA, GATE, Hiedel Time, MALLET, NegEX, SUTime) and 11 clinical (e.g., MEDLEE, TextHunter, v3NLP, cTAKES, ClinREAD, Clamp, MetaMap, RedCap, V3NLP) systems/toolkits. Techniques utilized under each system/toolkit were mostly rule-based NLP (n=14 techniques), followed by hybrid NLP (n=8 techniques) and ML algorithms embedded in NLP (n=4 techniques). The rule-based approach uses rules created by experts to categorize or label unstructured PRO data [71, 86], and some studies validated the rules with unstructured data from different patient samples and subsequently applied the rules to the new samples [87]. ML-based approaches often train established classification algorithms with statistical inference techniques through previously annotated text-corpus [13]. Once ML algorithms learn the unstructured data of a new system, they can be applied to other lexical, semantic, and contextual meanings without referring to any rules [88]. In contrast, the hybrid approach adopts both rule- and ML-based methods, and integrates expert-generated, ruled-based systems to guide ML algorithms to perform the analysis [16, 45, 58]. Fifteen NLP systems/toolkits were rule-based (e.g., generic toolkits: ConText, FMA, Heidel Time, NegEX, Protégé, SUTime, and Wordnet; clinical toolkits: CliX NLP, ClinREAD, MedLEE, MedEx, NLP-PAC, and V3NLP). Specific ML-embedded systems/toolkits included MALLET and GENSIM with R, and hybrid systems/toolkits included GATE, UIMA, Clamp, cTAKES, MetaMap, and MedTagger. A full list of references is in Supplementary Table S4.

Table 3:

The reported NLP systems or toolkits^*

Systems/toolkits	Full Names	Purposes**	Techniques***
Generic toolkits
ConText	N/A	Feature extraction and representation: sentence classification	Rule-based NLP
FMA	Freetext Matching Algorithm (FMA)	Preprocessing: annotation Feature extraction and representation: information extraction	Rule-based NLP
GATE	General Architecture for Text Engineering	Preprocessing: tokenization, sentence splitting, POS tagging, annotation Feature extraction and representation: name entity recognition, information extraction Non-neural ML: sentence classification	Hybrid NLP: rule-based, non-neural ML (support vector machine; SVM, WEKA ML)
GENSIM (R)	GENSIM	Feature extraction and representation: topic modeling and word embedding	Non-neural ML
Heidel Time	High quality rule-based extraction and normalization of temporal expressions	Preprocessing: tagging, normalization Feature extraction and representation: information extraction	Rule-based NLP
MALLET	MAchine Learning for LanguagE Toolkit	Feature extraction and representation: document classification, clustering, topic modeling, information extraction	Non-neural ML
NegEX	N/A	Feature extraction and representation: negation, affirmation	Rule-based NLP
Punkt Sentence Tokenizer	nltk.tokenize.punkt module	Preprocessing: tokenization	Non-neural ML
Protégé		Feature extraction and representation: ontology editor or framework	Rule-based NLP
NLTK (Python)	Natural Language Toolkit	Preprocessing: text tokenization, stemming, stop word removal, classification, clustering, POS tagging, parsing, and semantic reasoning	Non-neural ML
SUTime	Stanford NLP annotator	Preprocessing: annotation, recognizing and normalizing time expressions (TIMEx)	Rule-based NLP
Tagger_Date Normalizer plugin	Not available	Preprocessing: tagging, normalization	Rule-based NLP
TextHunter	N/A	Preprocessing: tokenization, stemming, POS tagging Feature extraction and representation: information extraction Non-neural ML: automated concept identification	Hybrid NLP: rule-based, non-neural ML (SVM based “batch learning”)
WordNet	N/A	Lexical database for NLP (ontology)	Rule-based NLP
Clinical NLP toolkits
Clamp	Clinical Language Annotation, Modeling, and Processing Toolkit	Preprocessing: tokenization, POS tagging, annotation, sentence boundary detection, section header identification Feature extraction and representation: assertion, negation, name entity recognition, UMLS encoder Non-neural ML: sentence boundary detetion, section header identification, classification	Hybrid NLP: rule-based NLP, non-neural ML (conditional random fields, CRF).
cTAKES	clinical Text Analysis and Knowledge Extraction System	Preprocessing: sentence boundary detection, tokenization, parsing, dictionary lookup annotation, normalization, POS tagging Feature extraction and representation: affirmation, negationname d section identification, name entity recognition, information extraction Non-neural ML: classification of medical information	Hybrid NLP: rule-based NLP, non-neural ML (CRF, SVM)
CliX NLP	Clinical NLP tools for SNOMED CT	Feature extraction and representation: processing system based on SNOMEDCT	Rule-based NLP
ClinREAD	Rapid Clinical Note Mining for New Languages	Preprocessing: tokenization, POS tagging, vocabulary mapping Feature extraction and representation: name entity recognition	Rule-based NLP
MedLEE	Medical Language Extraction and Encoding System Processing System	Preprocessing: parsing Feature extraction: clinical entities extraction, assertions Non-neural ML: word disambiguation, classification of medical information, generate rules for classifying medical conditions	Hybrid NLP: rule-based NLP, non-neural ML (CRF, SVM)
MedTagger	N/A	Preprocessing: tokenization, POS tagging Feature extraction and representation: information extraction, assertion, negation Non-neural ML: sentence detection, concept identification, patient level risk factor classification	Hybrid NLP: rule-based NLP, non-neural ML (WEKA ML)
MetaMap	N/A	Preprocessing: vocabulary mapping, parsing, tokenization, POS tagging, sentence boundary determination Feature extraction and representation: lexical lookup of input words in the SPECIALIST lexicon – an information extraction system based on UMLS	Rule-based NLP
MTERMS	Medical Text Extraction, Reasoning and Mapping System	Preprocessing: parsing, tokenization, POS tagging, vocabulary mapping, information extraction Feature extraction and representation: affirmation, negation	Rule-based NLP
MedEx	Medication Information Extraction System for Clinical Narratives	Preprocessing: tokenizer, tagging, semantic tagger, parsing, encoding Feature extraction and representation: information extraction	Rule-based NLP
NLP-PAC	NLP algorithms for Predetermined Asthma Criteria	Feature extraction and representation: information extraction, affirmation, negation	Rule-based NLP
V3NLP	Not available	Preprocessing: annotation Feature extraction and representation: information extraction	Rule-based NLP

^*See Supplementary Table S4 for a list of references;

^**The purpose of NLP systems/toolkits used to process unstructured PROs;

^***Hybrid NLP approach uses both rule-based and ML-based methods

NLP/ML Techniques for Processing Unstructured PROs

Table 4 reports the 3-step NLP/ML methods to process unstructured PROs from EHR clinical narratives: Step 1 is data pre-processing (n=60 studies), Step 2 is feature extraction and representations (n=69 studies), and Step 3 is data analysis (n=61 studies, including 39 using non-neural NLP/ML methods and 22 using neural NLP/ML methods). The step-by-step process was also summarized in Figure 2. The most common techniques for data pre-processing were annotation and text tokenization. The most common techniques for feature extraction and representations were rule-based NLP, affirmation/negation, and word2vec/bag-of-words (BOW). Regarding NLP/ML analytic methods, the most common non-neural ML methods were Support Vector Machine (SVM), decision tree, and Conditional Random Field (CRF). In contrast, commonly used neural NLP/ML methods included Convolutional Neural Network (CNN), Recurrent Neural Network (RNN; e.g., LSTM), and Artificial Neural Network (ANN or feed forward network). As an example of the way neural NLP/ML methods are used in clinical settings, the novel contextual embeddings of the BERT model use a context-based representation of chief complaints to predict specific signs/symptoms (e.g., pain, cognitive confusion) labeled by the experts and map semantically similar chief complaints to nearby points of vector space [7]. A full list of references was provided in Supplementary Table S5.

Table 4:

The 3-step NLP/ML application and corresponding techniques among the 79 studies included in the systematic review^*

Steps and techniques	N	%
Step 1: Preprocessing	60	75.9
Annotation	38	48.1
Text tokenization	36	45.6
Remove stop-words	18	22.8
Part-of-speech (POS) tagging	16	20.3
Normalization	14	17.7
Lemmatization/ stemming	12	15.2
Step 2: Feature extraction and representations	69	87.3
Rule-based NLP	37	46.8
Affirmation/ negation	33	41.8
Word2vec/ bag-of-words (BOW)	23	29.1
Name entity recognition (NER)	16	20.3
N-gram (Term Frequency–Inverse Document Frequency [TF-IDF], Document-Term Matrix [DTM], Term-Document Matrix [TDM])	15	19.0
Latent Dirichlet Allocation (LDA) for topic modeling	5	6.3
Latent semantic indexing (LSI)	1	1.3
Knowledge graph	1	1.3
Step 3: Data analysis (non-neural ML)	39	49.4
Support vector machine (SVM)	18	22.8
Decision tree (DT)	6	7.6
Conditional random fields (CRF)	9	11.4
Logistic regression classifier	8	10.1
Naïve Bayesian	6	7.6
Random forest (RF)	6	7.6
K-means clustering	3	3.8
K-nearest neighborhood (KNN)	3	3.8
Boosting (e.g., Light Gradient Boosting Machine [LightGBM], eXtreme Gradient Boosting [XGBoost])	2	2.5
Linear regression classifier	2	2.5
Bagging	1	1.3
Step 3: Data analysis (neural ML)	22	27.8
Convolutional neural network (CNN)	10	12.7
Recurrent Neural Network (RNN) (e.g., Bi-LSTM, GRU, Glove)	10	12.7
Artificial neural network (ANN) Feed forward network (FFN)	7	8.9
Transformer (e.g, Bidirectional Encoder Representations from Transformers [BERT], Bio-BERT)	3	3.8
Auto encoder	3	3.8
Embeddings from Language Model (ELMo)	1	1.3
Others	2	2.5

Abbreviations: Bi-LSTM, Bi-Long Short-Term Memory; BERT, Bidirectional Encoder Representations from Transformers.

^*See Supplementary Table S5 for a list of references

Figure 2:

NLP/ ML pipeline for processing unstructured PRO data in EHRs

DISCUSSION

This systematic review study focuses on PRO-related studies and the findings have a significant contribution to the literature by summarizing the NLP/ML applications in PROs (e.g., classification, phenotyping, and clustering; PRO-based risk prediction and stratification for clinical outcomes), the available NLP systems or toolkits, and NLP/ML pipelines (e.g., preprocessing, feature extraction and representations, and data analysis using non-neural or neural ML methods). Among 79 selected studies, most studies (>90%) used NLP/ML techniques for extracting free-text PRO data, followed by predicting the risk of adverse events (30%), classifying, phenotyping or clustering PROs (20%), and testing associations between PROs and clinical outcomes (8%). Given the challenges of using standard surveys to assess PROs in busy clinics, as well as well-described barriers to PRO instrument application into routine clinical care [2], NLP/ML application provides a convenient mechanism to integrate PROs available in EHRs into clinical workflows for clinical decision-making [89].

We found that different types/units of unstructured PROs (e.g., keyword/phrase, sentence, paragraph, entire document/note) were used in NLP/ML analyses. Vocabularies from standard clinical terminologies or nomenclatures (e.g., SNOMED-CT, UMLS, LOINC, DSM-5, and ICD-codes) were commonly used to process unstructured PROs. These rule-based systems (e.g., ontologies, medical terminologies) were typically used to identify the meaning of the words and terms from free-text PROs [90]. Systematized Nomenclature of Medicine–Clinical Terms (SNOMED–CT) is deemed the most comprehensive computer collection of medical terms and medical relationships [91]. The feasibility of mapping other medical terminologies (e.g., ICD-9 or ICD-10 codes) to the SNOMED–CT makes the translation between different terminology systems feasible. To achieve semantic interoperability [90]. many studies used medical vocabulary systems to map PRO words or terminologies [13, 17, 25, 28, 36, 37, 40, 44, 53, 57, 62, 76, 78]. Additionally, to account for disease-specific content, some studies created dictionaries for incorporating existing clinical terminologies and/or added new terminologies to complement the functionality of extant rule-based or vocabulary systems [12, 14, 16, 28–31, 37–40, 55, 60, 64, 69, 72]. In the clinical setting, PRO data are often created by clinicians through handwriting or typing in an unstructured format and stored in EHRs, and then coders follow medical terminologies (e.g., ICD diagnosis code, SNOMED-CT, UMLS) or vocabulary systems to map or annotate the unstructured data into the structured format. Our review study details the implementation of different NLP/ML techniques for transforming unstructured PROs into vectorized structured formats, together with the ML techniques for clinical use (e.g., disease prediction).

The studies included in our review used 15 generic and 11 clinical NLP systems/toolkits to process unstructured PRO data and transformed information into structured PROs. NLP systems/toolkits typically transfer unstructured PROs into numerically computable information through pre-processing free-text PRO and clinical data, extracting specific features of the data, and then data normalization (i.e., converting a token into its base form). The systems/toolkits for free-text PRO processing included rule-based NLP, ML-based, or hybrid approaches. Practically, NLP/ML techniques were commonly used for feature extractions and representation (n=70), and nearly half of these studies (n=37) relied on rule-based methods [6, 16, 17, 20–22, 25–28, 30, 31, 33–36, 39, 45, 49–51, 55, 56, 58–60, 64, 70–72, 74, 76, 80, 81, 83–85]. However, the rule-based method is not efficient because it requires manual extraction of knowledge that may involve trial- and-error [92]. Furthermore, the accuracy of the rule-based model depends on appropriate and available rules and domain expertise. Few studies included in our present review found comparable or superior performance of ML-based or hybrid approaches (e.g., accuracy, precision, recall, F1) to that of the rule-based NLP approach [51, 53, 74, 75]. Technically, ML-based or hybrid approaches learn the patterns between phrases and sentences, which improves accuracy and generalizability in representing clinical narratives beyond rule-based approaches [16, 45, 58].

Modern neural network-based ML algorithms (e.g., RNN, transformer-style model) are a breakthrough for processing unstructured clinical narratives. Several studies in our review reported better performance of newer neural network-based ML algorithms (e.g., BERT, GPT-3) compared to traditional NLP or other neural network-based ML algorithms (e.g., ELMO, LSTM) [7, 24, 53]. The superior performance of the BERT model to other NLP/ML methods is because the BERT model uses effective pre-training methods (i.e., masked language modeling and next sentence prediction) and deploys multiple transformer layers to account for contextual information of natural language [93]. The BERT model has been shown to effectively capture linguistic features (i.e., syntactic, semantic, contextual features) [94]. In comparison, the traditional feature extraction and representation methods (e.g., bag-of-words, NER, n-gram) require domain knowledge to design complex feature engineering with less generalizability [95]. As a result, the semantic features derived from the BERT model may incorporate more contextualized meanings of words than traditional NLP models when contexts of words vary. Recently, clinical BERT models (e.g., Bio-BERT, Cancer-BERT, and BlueBERT) have been developed, which continue fine-tune BERT models through the corpus of PubMed, MIMIC, or other clinical sources, to meet different purposes [96]. These findings suggest the usefulness of domain-specific neural network-based ML methods for processing unstructured PROs in the future.

While this review collects studies prior to 2021, we have covered the state-of-the-art NLP models (e.g., RNN, Bio-BERT, and Cancer-BERT) and those techniques have not been changed in PRO-related studies since 2020. The recent rise of deep learning approaches typically brings novel embedding techniques that can jointly integrate lexical, syntactic, semantic, and contextual patterns into unified vectors, which significantly reduces labor expenses and domain knowledge requirements of feature engineering and design. The core idea of developing embedding techniques is to use context to define language and therefore train embedding models (i.e., neural networks). Co-occurrence, word prediction, and neighboring sentence matching are useful approaches [94, 97] to obtain the embedding models for PRO research. Studies building deep learning models [6, 14, 50, 53, 55, 62] for PRO assessment benefit from the power and flexibility of embedding techniques to represent features from token to document levels [7, 24, 50, 53, 96]. Those studies have demonstrated their success in embedding techniques to learn representative features of free-text data in PRO research, such as phenotype inference [50], information extraction [14], and diagnosis [53]. Nevertheless, ML models trained on word embeddings output tables of weights can make debugging missed predictions more challenging compared to the models trained on simpler linguistic features.

Though various NLP systems/toolkits used to analyze unstructured PROs from clinical narratives have been reported [22, 27, 45, 48, 67, 81], the validity of using unstructured PROs derived from clinical narratives versus standard PRO surveys remains unclear. The standard PRO surveys may likely measure generic PRO concepts across all patients, whereas unstructured PROs from clinical narratives can capture patient-unique PROs that may not be included in the standard PRO surveys. Several studies in our review used PRO data collected from the survey to evaluate the validity of NLP/ML pipelines for unstructured PROs [22, 43, 45, 67]. Agreement of PRO data between data collected from self-reports and EHR-based unstructured PROs analyzed with NLP/ML ranged from a range of 67–82% [45]. A similar finding (63–75% agreement) was found in the patient responses to standard surveys versus unstructured symptom data (e.g., chest pain, dyspnea, cough) documented in the physicians’ medical notes [22]. Several studies have noted the value of unstructured PROs as anchors in predicting or correlating various clinical outcomes (e.g., disease onset, suicidal ideation, readmission, mortality) [20, 30, 33, 59, 61, 68, 79, 83]. These findings suggest that unstructured PROs may be a surrogate for the standard PRO surveys [93].

The findings of this review study have important implications for future clinical research and care. Although NLP/ML techniques are evolving, successful NLP/ML application requires the implementation of integrated platforms that seamlessly interconnect EHR functionality and NLP/ML algorithms to facilitate clinical interpretation. Evidence shows that integrating unstructured PROs into EHRs likely improves cancer survivorship care by predicting late effects based on worsening symptoms and other clinical data [7, 21, 30, 33, 44, 61, 63, 75, 79, 84]. However, there is a critical need to find novel platforms/systems for integrating unstructured PROs into EHRs and using NLP/ML techniques to annotate unstructured PROs. Further effort is warranted to create meaningful PRO scores from free-text narratives based on the features derived from NLP/ML to be comparably interpretable to scores from standard PRO surveys.

There are implementation barriers to collecting and integrating unstructured PROs from EHRs for NLP/ML applications. The major barriers include technical complexity, system interoperability, and concerns about the quality of unstructured clinical narratives (e.g., fragmented or incomplete words or sentences to represent a patient’s health status or symptoms) in EHRs [98]. Adopting the Common Data Model or Common Data Warehouse to enhance model portability [99] may improve the collection of standard and high-quality unstructured PRO data from clinical narratives. Instead of relying on text-based PRO data, alternative methods include the use of speech or voice recognition systems for collecting unstructured PROs and developing audio-based NLP/ML pipelines to automatically analyze PRO information from patient-doctor conversations in clinics. There are different ways to manage conversation-based PROs: 1) using NLP/ML algorithms to automatically annotate and analyze the conversation-based PROs and 2) using software to transcribe voice data, followed by NLP/ML to annotate and analyze the transcribed data. If systems were set up appropriately, acoustic recordings of clinical interactions would involve less paperwork and be less prone to data collection artifacts. However, compared to text, audio contains richer information (e.g., prosody) that humans process together with lexical meaning to alter language’s propositional content. For example, an ironic phrasing of “yeah, that medication worked” captured in audio would be interpreted as a declaration of efficacy if it were simply converted into text. Once the voice-based PRO data are annotated, the NLP/ML pipelines could be applied for data analysis and clinical application. Few-shot learning, or prompting, could also be used to improve parts of PRO prediction by significantly augmenting the amount of relevant ground truth training data [100].

This review study contains some limitations. First, several selected studies did not report specific information to meet our inclusion criteria (e.g., frequently used linguistic features); therefore, the results may not fully reflect the status of NLP/ML applications in unstructured PROs. Second, we did not evaluate the quality of the selected studies because the standards or guidelines for evaluating NLP/ML applications in unstructured PROs have yet been established. Finally, we included studies that were published by the end of 2020, while NLP/ML techniques evolve day by day. However, to our best knowledge, our study has covered the latest trends and techniques of applying NLP/ML techniques (e.g., BERT) for researching unstructured PROs. We are aware of large generative language models (e.g., ChatGPT), however, those newer generative AI techniques have not been applied in recent PRO-related studies.

While the NLP revolution, specifically generative AI, has taken place in the past 2–3 years, our literature search until 2023 found no instances of generative AI being applied to unstructured PROs. Large language models become the promising direction for future PROs studies, and contemporary NLP techniques (e.g., BERT or GPT) discussed in this review study can provide a solid background for future PROs research. One notable example of generative AI is ChatGPT (e.g., GPT-4 developed by OpenAI) which aims to generate a patient summary using information provided by clinicians or patients and to facilitate interactive text-based communication with users. It is important to note that although ChatGPT can address potential bias in the training data, it has not undergone a comprehensive fine-tuning and validation process for PROs and other medical data. Moreover, its integration into EHRs for annotating and analyzing vast amounts of unstructured PROs in medical notes remains uncertain [101–103]. Currently, ChatGPT is valuable for providing a PRO summary based on the data provided by clinicians or patients, but it should not be relied upon for clinical interpretation or decision-making until empirical studies establish its validity. We believe the methods and resources of traditional NLP/ML reported in this review will be valuable to inform the generative NLP approaches (e.g., GPT) of how to incorporate domain knowledge and promote evidence-based PRO inferences from unstructured medical data. Our present recommendation is to employ established, validated NLP/ML approaches for feature extraction, including rule-based or neural machine/deep learning algorithms, and utilize NLP systems/toolkits for the analysis of unstructured PROs within EHRs.

CONCLUSION

This systematic review study reports the usefulness of NLP/ML techniques in processing unstructured PRO data. Currently, using the established rules to annotate unstructured PROs through NLP/ML systems/toolkits is the dominant method, though the use of novel neural ML-based methods is increasing. Transformer NLP/ML models (e.g., BERT) are the most cutting-edge and dominating techniques to process unstructured PROs in EHRs. Although we did not come across any studies utilizing generative large language models like ChatGPT for extracting and analyzing unstructured PROs and integrated into EHRs, the rapid advancements in generative AI offer a potential opportunity for future exploration and evaluation in this area.

Highlights.

• Most of the studies used NLP/ML techniques to extract PROs from clinical narratives (81%), map the extracted PROs into specific PRO domains for phenotyping or clustering purposes (33%), process PROs for predicting disease progression or onset of adverse events (28%) or develop/validate NLP/ML pipelines for analyzing unstructured PROs (24%).

• Using annotation rules for NLP/ML to analyze unstructured PROs is dominant, yet deploying novel neural ML-based methods is warranted.

DATA AVAILABILITY

Extracted data for this systematic review is available upon request.