Capturing Individual-level Social Determinants from Clinical Text

Abstract

Knowledge of social determinants of health (SDOH), which refer to nonmedical factors influencing health outcomes, can help providers improve patient care. However, SDOH are often documented in unstructured notes, making them more inaccessible. Although previous works have attempted SDOH extraction from clinical notes, most efforts defined SDOH more narrowly and focused on the note’s social history (SH) section, where social factors are traditionally documented. Here, we introduce a new SDOH dataset covering a broad range of SDOH content that is annotated over entire notes. We characterize what, where, and how SDOH information is documented in clinical text, present baseline systems using a token classification and generative approach, and investigate whether training only on the SH section can effectively extract SDOH from the entire note. The final dataset, consisting of 2,007 annotations covering 7 open-ended SDOH domains over 500 notes, will be publicly released to encourage further research in this area.

Introduction

Social determinants of health (SDOH), defined by the World Health Organization as “the conditions in which people are born, grow, work, live, and age,”[1] have been shown to significantly impact an individual’s health, well-being, and quality of life[2]. Traditionally, the discourse on SDOH has revolved around health policy, health disparities, and public health, i.e., community-level determinants[3, 4, 5]. However, for providers at the point-of-care, individual- level determinants are likely of more interest because they are immediately actionable[5]. To address a patient’s social needs, providers must first be aware of them. Although SDOH information is commonly captured as part of the patient history during clinical encounters and documented accordingly in clinical text, including details of the patient’s family makeup, occupation, substance use, etc.[6], the unstructured nature of clinical documentation makes it difficult for clinicians to retrieve this information later. Add to this the complexity introduced by multiple providers caring for a single patient and the dynamic nature of individual-level determinants[5] (e.g., a patient with a health insurance lapse due to job loss may no longer have similar needs after finding a new job), and it becomes increasingly difficult for providers to obtain a complete picture of the patient’s social determinants.

To improve provider awareness of individual-level determinants, we envision a timeline view of SDOH information that allows users to visualize changes in a patient’s social factors over time. The use of timelines to visualize a longitudinal history of the patient has been well-documented[7, 8, 9, 10, 11]. To explore this idea, we conducted a preliminary review of clinical notes within the 2014 i2b2 shared task corpus[12, 13], as it contains longitudinal data of 2-5 notes per patient and uncovered examples of how social factors may change over time for a given patient. Figure 1 shows one such example derived from a series of notes for a single patient in the 2014 i2b2 shared task corpus, where, over the span of six years, the notes document a change in health insurance, followed by dissatisfaction with the new insurance provided by his employer, to job loss with loss in insurance and accompanying legal issues leading to increased stress, anxiety and a relapse in smoking. Two observations can be gleaned from such an example: (1) patient-level determinants are dynamic and can change over time, and (2) social factors are interrelated and should be considered holistically to provide a complete picture of the patient’s social situation, e.g., the interplay between health insurance, economic stability, stress, and smoking behavior shown in Figure 1.

Figure 1:

Series of notes for a single patient over 6 years, demonstrating the interplay among various social determi- nants of health and the importance of considering these factors holistically.

Prior research has demonstrated the use of natural language processing (NLP) techniques to extract SDOH informa- tion from clinical text[14]. However, most work has been conducted on proprietary datasets unavailable to the general research community[15, 16, 17, 18]. Of the publicly available SDOH datasets we reviewed, some focus on a smaller subset of determinants depending on the specific use case, such as substance use[19, 20], while others cover a larger variety of determinants but still tend to define them in very narrow terms. For example, MIMIC-SBDH[21] only captures the patient’s housing status (patient has housing, is homeless, or no information) as the environment determinant. However, topics like quality of housing, access to transportation, availability of healthy foods, air and water quality, and neighborhood crime, are also aspects of the environment that can affect an individual’s personal health[2].

Recently, the National NLP Clinical Challenges (n2c2) organized a shared task for extracting SDOH using the Social History Annotation Corpus (SHAC)[22, 23]. Although SHAC is a relatively large corpus with detailed annotations for 12 SDOH, its overall coverage of the SDOH area is still somewhat limited. For example, SHAC does not directly capture any economic instability and financial strain experienced by the patient. Within SHAC, such information may be inferred through its annotations for employment status (if unemployed, retired, or on disability), insurance status (if no insurance), and living status (if homeless). However, an individual can be employed and have insurance and housing but still struggle financially to afford day-to-day expenses.

In creating their datasets, both MIMIC-SBDH[21] and SHAC[23] extracted the social history section from the note and annotated SDOH only on the extracted social history sections. While information relevant to the patient’s social and behavioral risk factors is traditionally documented in the social history section, such information may also be woven into other parts of the clinical note, especially when relevant to the patient’s care and management, e.g., descriptions of the patient’s level of physical activity and dietary patterns as part of diabetes management, or cost considerations when deciding what drug to prescribe. In addition, some notes may simply not have a designated social history section. Considering the narrower scope of existing publicly available SDOH datasets and the interrelatedness of different SDOH topics in reality, there is a need for a dataset that encompasses a broader, more comprehensive range of SDOH information in clinical text.

In this work, we present the Descriptive Social determinants Corpus (DSC), a dataset of 500 clinical notes annotated with SDOH information, with coverage over a wider range of SDOH content beyond what previously released datasets contain. We built off of the five domains proposed by the Office of Disease Prevention and Health Promotion, Healthy People 2030[24]: Economic Stability, Education Access and Quality, Health Care Access and Quality, Neighborhood and Built Environment, and Social and Community Context. We then added two additional high-level domains, Behavioral and Other SDOH, to cover other determinants of health not covered by the previous five categories. In contrast to previous work, we adopt a more open-ended approach to SDOH extraction, capturing more nuanced descriptions of determinants not limited by rigid frameworks or schemas. Our objective is to develop a high-recall system able to extract a wide variety of SDOH from clinical text. The major contributions of this paper are:

• A publicly available dataset of 500 clinical notes annotated with 2,007 SDOH text snippets, and coverage over a broader range of SDOH content as compared to previously released datasets,

• Insights into what types of SDOH information are recorded in the clinical text (i.e., subtopics), how they are recorded, and where (i.e., standalone section vs woven into the clinical narrative),

• Experimental results showing that the problem can be successfully modeled as a token classification (i.e., Named Entity Recognition (NER)) or a generative task, the importance of using in-domain pre-trained large language models, and error analysis providing insights into the most challenging aspects of this task.

Methods

Data and annotation

DSC was created on the subset of clinical notes from the 2014 i2b2/ UTHealth Natural Language Processing shared task corpus[12, 13] used in the Contextualized Medication Event Dataset (CMED)[25], which contains 500 notes over 296 patients. The CMED corpus was used because of its note selection process that ensured a balance of longitudinal and heterogeneous data. Of the 500 notes annotated, 100 were doubly annotated and adjudicated to measure inter-annotator agreement. Notes were annotated by a team of two annotators led by a physician. Annotators were asked to review each note in full and annotate any SDOH information that may inform the patient’s medical care, with the understanding that the annotated spans would be presented independently to an end-user, i.e., each annotated span must be self-contained. In developing our annotation guidelines, we consulted descriptions of social determinants of health from Healthy People 2030[24], recommended domains from the Institute of Medicine report on capturing social and behavioral domains in electronic health records[3, 4], and annotations from previously released SDOH corpora[21, 23] to ensure we achieve a broad coverage of determinants in our dataset. We decided on 7 domains, as described in Table 1: EconomicStability, EducationAccessQuality, HealthCareAccessQuality, NeighborhoodBuiltEnvironment, SocialCommunityContext, Behavioral, and OtherSDOH.

Table 1.

Annotation schema and examples.

Domain	Description	Sample annotation
EconomicStability	Financial resources (income, cost of living, socioeconomic status) - includes key issues such as poverty, employment, food security, and housing stability	“retired” “on disability from work” “is a set designer at Columbia”
EducationAccess Quality	- includes key issues such as graduating from high school, enrollment in higher education, educational attainment in general, language and literacy, and early childhood education and development	“at about age 68, he went back to school to study sociology” “has a 12th grade education”
HealthCareAccess Quality	People’s access to and understanding of health services and their own health - includes key issues such as access to healthcare, access to primary care, health insurance coverage, and health literacy	“has been having difficulty getting appt d/t new ins” “her understanding of her disease process appears to be limited”
NeighborhoodBuilt Environment	Where a person lives or works - includes topics like quality of housing, access to transportation, availability of healthy foods, air and water quality, and neighborhood crime and violence	“is at a nursing home” “lives in Ohio” “doesn’t have a car so needs help getting groceries”
SocialCommunity Context	Relationships and interactions with family, friends, co-workers, and community members - includes topics like cohesion within a community, civic participation, discrimination, conditions in the workplace, and incarceration - marital status, family makeup, who lives at home, who helps with medical care/ activities of daily living, social ties causing stress (e.g., family medical issues, interpersonal relations), social ties providing support (e.g., involvement in club, church)	“widower” “has 2 daughters” “lives w/ 17yo son” “wife with memory lapses which worries him”
Behavioral	Behavioral determinants of health - includes topics like dietary patterns, physical activity, tobacco use and exposure, alcohol use, abuse of other substances, sexual practices, exposure to firearms, risk-taking behaviors (e.g., distracted driving, helmet use)	“TOB: quit 20y prior” “daily exercise routine of jumping rope and pushups” “Does not wear seat belts”
OtherSDOH	Other non-clinical characteristics that may affect patient’s health - includes topics like country of origin, race/ethnicity, sexual orien- tation, gender identity, active stressors, disabilities affecting activi- ties of daily living, recent travel, hobbies	“born and raised in rural Finland” “work has been stressful” “uses a cane for ambulation”

Allowing for the possibility of SDOH content residing outside the social history section, we made two decisions during the creation of this dataset. First, we opted to annotate the entire note instead of annotating only the social history section as in prior works. Second, in addition to annotating SDOH, we also asked annotators to label the start and end of any social history sections in the note, when present, to allow for further analysis and experiments. Social history sections were identified based on their formatting and content. In most cases, the social history section has a section header indicating its start (e.g., “Social History”, “SocHx”, “Habits”). However, in some notes, content may be organized by paragraphs or chunks of text separated by line breaks without explicit headers. In these cases, if the annotator understood a paragraph to be intended to document the patient’s social history based on its content and language, it was labeled as social history section even without an explicit section header.

System description

To automate the task of SDOH extraction, we experiment with two different problem formulations. In the first, we employ Bidirectional Encoder Representations from Transformer (BERT[26])-based language models in a token classification setup by adding BIO (Beginning, Inside, Outside) NER tags to each input sentence from the sentence segmented clinical note. We experiment with models pre-trained on the general domain (BERT-base[26]) as well as in-domain (ClinicalBERT[27]) datasets. In the second setup, we formulate our problem as a generative task and employ pre-trained T5[28] models where the input is a sentence from the clinical note and the model is tasked to generate the SDOH tags inline. We experiment with the in-domain Clinical-T5-Base[29] models. The two different task setups with an input example and expected model output is presented in Figure 2. As there is considerable overlap between the SDOH domains (i.e., multi-label problem), we train a different model corresponding to each SDOH domain in both setups. We also experiment with clinical note-tuned sentence segmentation and tokenization, and a post-processing step involving stop word removal to improve precision. We summarize our experimental configurations as follows:

1. Token Classification Task - General domain (BERT-base) and in-domain (ClinicalBERT) pre-trained language models with pre and post-processing

2. Generative Task - Clinical-T5-Base pre-trained language model with pre and post-processing

Figure 2:

Task setup for token classification task using BERT (A) and generative task using T5 (B).

As we have chosen to use the CMED corpus due to its note selection process, we follow the same 75/5/20 splits for train/dev/test as specified by Mahajan et al. [25]. We also exclude HealthCareAccessQuality and EducationAc- cessQuality from our experiments because of their low ground truth instances (Table 2). We employ the pre-trained language model to obtain a distributed representation, dropout rate of 0.2, and a fully connected layer with softmax activation. We use the transformers package[30] to tune our models with the train and dev splits, and present our results on the test split.

Table 2.

Dataset statistics for our two experimental setups.

Domain	Count		Experiment
			Baseline			SH only vs full note
	Full note	SH only	Train	Dev	Test	Train (SH only)	Train (downsampled full note)
Behavioral	869	563	583	131	155	392	392
SocialCommunityContext	577	315	383	74	120	214	216
OtherSDOH	205	44	143	22	40	–	–
EconomicStability	185	155	131	23	31	110	107
NeighborhoodBuiltEnvironment	148	75	104	8	36	–	–
HealthCareAccessQuality	18	0	12	5	1	–	–
EducationAccessQuality	5	5	4	0	1	–	–
TOTAL #annotations	2007	1157	1360	263	384	716	715
TOTAL #sentences	25729	977	17808	2948	4973	641	9744

In another set of experiments, we evaluate whether a model trained on just the social history (SH) section of a note can effectively extract SDOH from the entire note. Since there are limited ground truth instances within the SH section for certain domains, for these experiments, we only explore domains with at least 100 ground truth instances within the SH section. In setting up our experiments, we keep the same train/dev/test split as before and compare 2 sets of domain-specific models (both based on the best configuration in our previous set of experiments): one set of models trained only on annotations contained within the social history section (SH only), and the other set trained on annotations from the full note. As the ground truth is significantly less for the SH only models, we downsample the ground truth for the full note models by removing patients from the train split until a comparable ground truth size is achieved. The dev and test datasets remain unchanged from our previous set of experiments, i.e., dev and test splits use annotations from the full note. Table 2 presents the number of annotations in each split for all experiments.

We evaluated our experiments on the 20% test data with lenient matching (i.e., any overlap is considered a true positive). We present precision, recall, and F₁ scores for each experiment in the Results section.

Results

Inter-annotator agreement and dataset statistics

After several training sessions, annotators achieved good inter-annotator agreement with a Cohen’s kappa of 0.79. The final annotated dataset consists of 2,007 SDOH annotations over 500 notes. Table 2 presents the distribution of annotations over specific domains. In addition to the SDOH annotations in Table 2, DSC also contains 286 SH Begin and End annotations which are present in 284 notes (some notes have more than one SH section).

As observed in Table 2, Behavioral and SocialCommunityContext account for the majority of the annotations at 43.3% and 28.7%, respectively, while HealthCareAccessQuality and EducationAccessQuality have the least number of annotations (<1%). Also of note, a substantial percentage of annotations (42.4%) was found outside the social history section, including almost half of SocialCommunityContext (45.4%) and NeighborhoodBuiltEnvironment (49.3%), and more than half of OtherSDOH (78.5%) and HealthCareAccessQuality (100%), demonstrating the need for SDOH extraction beyond the social history section.

System results

Results for our two experimental setups are presented in Table 3. BERT-base provides a competitive baseline on all five SDOH domains. Further incorporating in-domain pretraining by employing ClinicalBERT improves the performance on most of the domains. For the generative approach, Clinical-T5-Base provides a lower or comparable performance to ClinicalBERT. As our dataset contains longer phrases compared to an average NER task, we hypothesize that in the token classification setup, the model is able to learn well as the neighboring tokens are also tagged similarly as compared to the text-to-text generative task where the model may have a more challenging time capturing the entire longer phrase with a single label at the end. Pre and post-processing modules further boost the performance of both models, especially for ClinicalBERT, making ClinicalBERT + Pre&Postprocessing our best-performing system.

For the SH only vs. full note experiments, the full note models performed significantly better compared to the SH only models, as shown in Table 4. SH only models achieved high recall but low precision, which can likely be attributed to a low number of negative examples in the SH only training data (i.e., low number of sentences as shown in Table 2).

Table 4.

Results of models trained on annotations in social history (SH) only vs full note. Bold indicates the highest precision, recall, and F₁ scores for each domain.

Domain	SH only			Downsampled full note
	Precision	Recall	F 1	Precision	Recall	F 1
Behavioral	0.37	0.88	0.52	0.88	0.85	0.87
SocialCommunityContext	0.44	0.83	0.57	0.77	0.81	0.79
EconomicStability	0.27	1.0	0.43	0.71	0.88	0.77

Discussion

Dataset analysis

In developing our annotation guidelines, we purposefully defined our 7 domains in very open-ended terms to capture a wide variety of determinants within our framework. To better understand the resulting dataset, we analyzed annotations for each domain to better characterize our dataset and gain insights into the types of SDOH captured within DSC. Figure 3 shows a breakdown of the 5 major domains (by ground truth count) into more granular topics of interest.

Figure 3:

Distribution of subtopics for 5 major domains by ground truth counts, which are presented next to the corresponding subtopic label. Counts of <15 are not shown.

Here we highlight some specific types of SDOH information that are unique to this dataset. As observed in Figure 3, although the majority of Behavioral annotations are regarding substance use (73.6%), i.e., tobacco, alcohol, and drug use, this dataset also has a fair amount of annotations for physical activity (13.8%) and dietary behaviors (9.6%). As for SocialCommunityContext, annotations included not only cover the common measures usually captured (e.g., marital status, family composition, who lives at home), but also specific statements describing the type of social support (e.g., “Daughter helps her do errands”, “accompanied by his son today”) as well as descriptions of stress resulting from social ties that may be affecting the patient’s health and well-being (e.g., “is extraordinarily upset about her daughter who is homeless”). OtherSDOH is the most diverse domain, as it was designed as a catch-all for any non-clinical characteristics affecting the patient’s health not captured by the other 6 defined domains. In practice, approximately a third of OtherSDOH (36.6%) described issues with disabilities affecting activities of daily living (e.g., “wheelchair bound”, “legally blind”). Another third (34.6%) described the patient’s ethnicity or place of origin (e.g., “caucasian”, “originally from Hawley, Puerto Rico”). The remaining third (28.8%) is the most interesting with the richest variability in expression, including information such as recent travel (e.g., “recently in Amboy”), language needs (e.g., “speaks Spanish and broken English”), healthcare concerns/preferences (e.g., “pt refuses mammogram. She continues to be suspicious that screening tests are experimental”), worries/stresses (e.g., “worried about having to be placed in nursing home because cannot do own med preparation”), etc.

In a separate analysis, we reviewed what types of information (i.e., subtopics) are contained inside vs. outside the social history section for each domain. Interestingly, although most domains have over half of their annotations within the SH section (Table 2), most of the annotations within the SH section belong to the dominant subtopic within that domain. In other words, information relevant to the less common subtopics are mainly found outside of the SH section. Therefore, if one were to look only at the SH section without considering the entire note, one would be more likely to miss out on less common subtopics. For example, within Behavioral determinants, although the majority of substance use annotations are found within the SH section (76.9%), most mentions of physical activity and diet are found outside of the SH section (69.1% and 78.3%, respectively). Figure 4 illustrates this observation using word clouds¹ to compare the information inside vs. outside the SH section for Behavioral determinants. The 5 most common words for Behavioral determinants inside the SH section (i.e., “etoh”, “alcohol”, “denies”, “tobacco”, “smoking”) are all related to substance use, while the 5 most common words for Behavioral determinants outside the SH section (i.e., “smoking”, “diet”, “smoker”, “exercise”, “day”), include terms related to physical activity and diet. Similarly, for SocialCommunityContext, although annotations for family composition and living situation are mostly found within the SH section (71.1% and 90.8%, respectively), annotations for other subtopics are mainly found outside of the SH section (85.5%, 77.5%, and 73.3% for social support, stress, and VNA, respectively). The differences in the types of information documented within and outside the social history section further highlight the need to consider the full note when extracting SDOH.

Figure 4:

Comparison of the 40 most frequent words in annotated Behavioral determinants found inside vs outside the social history section. Words inside social history mainly relate to substance use (e.g., “etoh”, “alcohol”, “smoking”, “tobacco”) as compared to words outside social history that include more terms associated with physical activity and dietary behaviors (e.g., “diet”, “exercise”, “walking”, “active”, “eats”).

Error analysis

We conducted error analysis on the best-performing baseline model (i.e., ClinicalBERT + Pre&Postprocessing). We identified 4 major categories of errors leading to false positives (FPs) resulting in low precision: (1) not considering context, (2) annotation guidelines, (3) proper nouns, and (4) attributable to another individual. Errors due to not considering context affected all domains. For example, “single” in “single-level apt” was incorrectly predicted as SocialCommunityContext. Similarly, “Former” in “Former tobacco user” was predicted as EconomicStability, likely due to ground truth annotations that have a similar phrasing (e.g., “former banking manager”.) In the case of errors due to annotation guidelines, most FPs can be attributed to a specific guideline for SocialCommunityContext, namely, to only annotate family members when it is documented in a way that indicates they have some active relationship with the patient. This nuance proved difficult for the model to learn, resulting in mentions of family members in the family history section of the note (which documents any diseases or health conditions in the patient’s family) being incorrectly predicted as SocialCommunityContext. Errors due to proper nouns mainly affected SocialCommunityContext and EconomicStability. One hypothesis for this behavior is that proper nouns, with the expected capitalization, are frequently found in the ground truth for SocialCommunityContext (e.g., name of family member) and EconomicStability (e.g., name of company). Therefore, incidental mentions that have similar capitalization (e.g., physician name) are incorrectly predicted by the model for these two domains. Within a note, sometimes information about other individuals may be documented incidentally, such as the patient’s spouse’s health or occupation. In these cases, the model sometimes incorrectly attributes that information to the patient. For example, “disabled” in “She spends a lot of time caring for her partially disabled husband” was incorrectly attributed to the patient and predicted as OtherSDOH. One way to address these issues is to introduce global information to contextualize predictions (e.g. note section information).

As compared to the other domains, models for NeighborhoodBuiltEnvironment and OtherSDOH had significantly lower recall. This can likely be attributed to insufficient ground truth, especially when we look at the ground truth counts for the subtopics within those domains (Figure 3) and consider the subtopics for the false negatives missed by the models. For NeighborhoodBuiltEnvironment, the model missed more uncommon expressions of housing lo- cation (e.g., “Florida, where pt. lives”) or rarer subtopics within this domain (e.g., “has 2 flights of stairs at home” - considered as Other in Figure 3). OtherSDOH is the most diverse domain, both in terms of the types of information included (i.e., subtopics) and in the way such information is expressed, i.e., tends to be longer, more descriptive, with more language variability. Therefore, as expected, the lack of sufficient ground truth for OtherSDOH had a substantial negative impact on overall model performance. Future efforts to expand the current ground truth are necessary to further develop these rarer domains and subtopics for training or fine-tuning models.

Results for the SH only vs. full note experiments showed that SH only models performed significantly worse than full note models (Table 4), mainly due to low precision. Low precision here is most likely due to the lack of negative examples during training, resulting in an amplification of the same types of FP errors observed in the ClinicalBERT models. For example, for Behavioral determinants, there were 56 FPs for the SH only model that contained terms related to timing or frequency (e.g., “history of “, “1-2x/year”, “couple of days per week”, “only 1-2 day”) due to not considering context. In these cases, the timing information was not associated with any behavioral activities (e.g., substance use, exercise), but instead associated with other information such as medication use or illness (e.g., “1-2x/year” in “Migraines occur 1-2x/year” incorrectly predicted as Behavioral).

Limitations

We acknowledge certain limitations to our work. DSC is built on top of the corpus used in the 2014 i2b2/UTHealth Natural Language Processing shared task [31]. This corpus focuses heavily on diabetes and heart disease patients; it is not representative of a typical patient population. Further, the corpus is limited to a single data warehouse. Reproduction of our work on more diverse corpora is needed to better understand the effectiveness and applicability of our schema. Regarding the modeling of the task, we have limited the model by providing only the sentence as the context. In the future, we aim to experiment with models using a larger context window to incorporate global information (e.g., note section information). Further, we have not experimented with the current zero-shot or few-shot state-of-the-art systems like ChatGPT² or GPT-4³ due to data-sharing limitations. As these models or their open-source versions become more available, we aim to evaluate their capabilities on the task of SDOH extraction. Finally, there may be a need for a post-processing step (e.g., extracting structured information such as time) to allow our current annotations to be used in downstream applications such as in clinical settings or in large population analyses. Future work can be undertaken to provide such information through an extraction task built on top of DSC.

Conclusions

We introduce DSC, a dataset capturing a broad range of SDOH content that is annotated over the entire note, consisting of 2,007 annotations covering 7 open-ended SDOH domains over 500 notes. We describe our annotation guidelines, discuss specific nuances observed during the annotation process, and explore token classification and generative text-to-text NLP techniques to automate the task. We also compare the performance of models trained only on the social history section vs. models trained on the full note, and emphasize the importance of an SDOH dataset that provides annotations over the entire clinical note. By making this dataset publicly available, our aim is to stimulate further investigation and inquiry into utilizing social determinants of health extracted from clinical narratives, while also making a valuable contribution to other use cases that require the consideration of SDOH events.

Figures & Tables

Table 3.

Evaluation of our various system configurations. We present lenient precision, recall, and F₁ scores for each SDOH domain. Bold indicates the highest precision, recall, and F₁ scores for each domain.

Experiment	Domain
	Behavioral			SocialCommunity Context			OtherSDOH			EconomicStability			NeighborhoodBuilt Environment
	P	R	F 1	P	R	F 1	P	R	F 1	P	R	F 1	P	R	F 1
BERT − base	0.65	0.74	0.70	0.70	0.87	0.78	0.45	0.40	0.42	0.62	0.90	0.71	0.73	0.60	0.67
ClinicalBERT	0.86	0.86	0.86	0.72	0.85	0.78	0.52	0.34	0.42	0.47	0.92	0.62	0.85	0.61	0.71
ClinicalBERT + Pre&Postprocessing	0.91	0.86	0.88	0.76	0.84	0.80	0.68	0.33	0.44	0.68	0.90	0.78	0.91	0.58	0.71
Clinical − T 5 − Base	0.89	0.61	0.72	0.92	0.58	0.71	0.88	0.18	0.29	0.71	0.87	0.78	0.74	0.39	0.51
Clinical−T 5−Base+ Pre&Postprocessing	0.89	0.61	0.72	0.91	0.59	0.72	0.88	0.18	0.29	0.71	0.87	0.78	0.75	0.42	0.54