Our primary data source is MIMIC-III [25], one of the largest publicly available ICU databases encompassing 12 years of data (2001-2012) from Beth Israel Deaconess Medical Center. First, we excluded admission data from patients who were less than 18 years old at the time of admission. For inclusion, admissions were also required to have at least one note from any of these 3 categories: discharge summary, social work note, or rehabilitation service note. We selected these 3 types of notes to maximize the use of social and behavioral information for SBDH extraction: Discharge summaries are a comprehensive summary of a patient’s hospital stay, social work notes focus specifically on the social nature of a patient’s life, and rehabilitation service notes focus on improving patients’ function and mobility to stabilize them for discharge. The final sample consisted of 48,869 hospital admissions from 37,361 patients. An overview of the data selection process is shown in Figure 1.

All baseline variables were grouped into 3 categories: demographic, clinical, and SBDH. The demographic variables included age (18-39 years, 40-64 years, >64 years), gender (male or female), race/ethnicity (White, Black, Hispanic, or others), and marital status (married, divorced, widowed, single, or unknown marital status). As clinical variables, we considered drug use disorder, bipolar disorder, tobacco use disorder, major depressive disorder, alcohol use disorder, cirrhosis, chronic obstructive pulmonary disease (COPD), and renal insufficiency. This comprehensive list was made based on earlier studies related to OD [26-29], clinical judgment, and statistical analyses (see the “Statistical Analysis” section for further details). All clinical variables were detected using the International Classification of Disease, 9th Revision (ICD-9) codes from the admission diagnosis chart and included as dichotomous variables. The list of ICD-9 codes is available in Multimedia Appendix 1.

For SBDH variables, we used NLP to analyze the unstructured text data available in MIMIC-III. For each type of note, we chose the most relevant sections to extract the SBDH information: (1) discharge summaries: “Social History” sections; (2) social worker notes: “Patient/Family Assessment,” “Past Addictions History,” “Past Medical History” sections; (3) rehabilitation services: “Sexual and Social History” section.

We used the popular clinical NLP tool medSpaCy [30] to extract these sections from a note. We randomly chose a note, extracted the relevant sections as mentioned, and annotated for 6 categories of SBDH information. This process was repeatedly followed until we reached 1000 notes with at least one SBDH annotation. This annotated subset was later used to train a Bidirectional Encoder Representations from Transformers (BERT) model to extract SBDH at the word level. BERT [31] is a state-of-the-art language representation model that has successfully outperformed many other NLP systems across a wide range of tasks. We used the trained model to predict the SBDH information for the remaining notes. For an admission with multiple notes of the same type, we took the last note as representative of that admission as it typically includes the content of all the previous notes.

The 6 SBDH variables we chose were (1) housing insecurity, (2) unemployment, (3) social isolation, (4) alcohol use, (5) tobacco use, and (6) illicit drug use. The first 3 are social determinants and were selected based on the list of well-accepted social determinants provided by the Kaiser Family Foundation [32]. The rest were substance use–related health risk behaviors (ie, behavioral determinants) that were chosen for their clinical significance and relevance to OD. Details about the annotation process, NLP model development, and SBDH variable extraction procedures are provided in Multimedia Appendix 2.

In addition to the NLP-derived SBDH variables, we also identified social determinants from the structured data. We used the ICD-9 codes from patient diagnoses [33] to construct these 3 SBDH variables: (1) housing insecurity, (2) unemployment, and (3) social isolation. These were later integrated with the NLP-derived SBDH variables and prioritized in case of any mismatch. For example, if the NLP system detected “housing insecurity” as “No” for an admission and we obtained “Yes” from that admission’s diagnoses codes, we considered “Yes” as the correct value. In the end, there were 41,669 admissions (41,669/48,869, 85.27%) with at least one SBDH variable. Table 1 illustrates the 6 SBDH variables with brief descriptions and examples. If an admission had no mention of SBDH information, SBDH variables were coded as “unknown.” For instance, if an admission had no mention of patient housing status in the corresponding notes, homelessness was considered “unknown.” Other than these 3 SBDH variables, we also extracted insurance provider (private, Medicaid, Medicare, other government, or self-pay) information using ICD-9 codes.

The outcome was nonfatal OD, which was identified using ICD-9 codes [34].

First, we performed correlation and collinearity analyses for all the variables. The correlation plot and variance inflation factor [35] did not show multicollinearity among the variables. For the clinical variables, based on earlier work and task relevance, we chose 14 comorbidities: posttraumatic stress disorder, major depressive disorder, bipolar disorder, schizophrenia, alcohol use disorder, drug use disorder, tobacco use disorder, hepatitis C, diabetes, congestive heart failure, obstructive sleep apnea, COPD, cirrhosis, and renal insufficiency. We built logistic regression models and employed the sequential forward selection procedure [36] to identify the most essential clinical variables related to OD. The final list included 8 clinical variables: drug use disorder, bipolar disorder, tobacco use disorder, major depressive disorder, alcohol use disorder, cirrhosis, COPD, and renal insufficiency.

We used a logistic regression model to examine the associations of nonfatal OD with demographic, SBDH, and clinical variables. This was assessed in terms of adjusted odds ratios (aOR) with 95% CIs. We also evaluated the crude odds ratio (OR) with 95% CIs. The statistical significance was measured at P<.05. Hosmer-Lemeshow test was conducted and indicated a sufficient fit for our model (χ₈=10.39; P=.24). All statistical analyses in this study were conducted in R (version 4.0.2).

Table 2 presents the characteristics of our cohort (n=48,869). Our sample was comprised of mostly men (27,436/48,869, 56.14%) and white (35,058/48,869, 71.74%) adults. The majority of patients were aged 64 years or older (25,276/48,869, 51.72%). Of the clinical variables, renal insufficiency was the most prevalent (8158/48,869, 16.69%), followed by COPD (5674/48,869, 11.61%) and alcohol use disorder (4121/48,869, 8.43%). In our cohort, we observed that 7.28% (3559/48,869) of the patients were unemployed, 13.35% (6523/48,869) were socially isolated, and 0.82% (402/48,869) had housing insecurity. We found 171 (171/48,869, 0.35%) admissions with nonfatal OD.

Of the 6 NLP-derived SBDH variables, only housing insecurity, unemployment, and social isolation had associated ICD-9 diagnostic codes. Compared with their NLP-derived counterparts, these structured variables were coded infrequently. For example, using ICD-9 codes, we found 258 admissions with “housing insecurity,” whereas the NLP system detected 402 admissions. For “unemployment,” it was 20 for the ICD-9 codes and 10,876 for the NLP system. And more striking, for “social isolation,” only 4 admissions had relevant ICD-9 codes in their diagnosis compared to 6523 admissions found by the NLP system. Due to the substantial prevalence gap, we did not compare the quality of these 2 types of SBDH variables side by side. In all, structured SBDH variables accounted for only 0.18% of the SBDH variables. This clearly shows that NLP can be useful to extract SBDH information from EHR notes when structured data are not enough. This also helps reduce bias from the use of structured data only.

Several factors were strongly associated with nonfatal OD (Table 3). Among the demographic risk factors, Blacks (aOR 0.51, 95% CI 0.28-0.94) and older age groups (40-64 years: aOR 0.65, 95% CI 0.44-0.96; >64 years: aOR 0.16, 95% CI 0.08-0.34) had lower odds compared with White and younger patients. Among the 8 clinical variables, 5 were strong risk factors for nonfatal OD. We observed increased odds of overdose among individuals with drug use disorder (aOR 8.17, 95% CI 5.44-12.27), bipolar disorder (aOR 2.69, 95% CI 1.68-4.29), and major depressive disorder (aOR 2.57, 95% CI 1.12-5.88). Interestingly, tobacco use disorder (aOR 0.53, 95% CI 0.32-0.89) and alcohol use disorder (aOR 0.64, 95% CI 0.42-1.00) had decreased odds. Among the SBDH variables, individuals with Medicaid had increased odds compared with those with private medical insurance (aOR 2.26, 95% CI 1.43-3.58). History of (aOR 2.09, 95% CI 1.15-3.79) and current (aOR 2.06, 95% CI 1.20-3.55) use of illicit drugs were also strongly associated with the outcome.

To our knowledge, this is the first study to examine the risk factors associated with nonfatal OD leading to ICU admission. In the United States, the need for characterizing critical care patients with OD is rising [23,24], and this study partially addressed that by identifying the risk factors for nonfatal OD from a large ICU database. The novelty also lies in the use of a state-of-the-art NLP system that utilized unstructured EHR notes for essential SBDH extraction due to inadequate representation from structured data. There is a growing body of literature showing that SBDH can strongly influence patient health and outcomes [12]. For example, SBDH variables have been shown to be strongly associated with suicide attempt [11], mortality [17], and mental health diagnosis [18]. The challenges here for the health care systems are to set up methods that can identify SBDH and use them at the point of care to inform clinical action [37,38]. Our work demonstrated that using NLP to detect SBDH information from EHR text can be a viable option in this regard.

According to our analysis, multiple SBDH variables were significantly associated with nonfatal OD in ICU settings. We observed that patients with economic instability (unemployed) were more likely to have an overdose, but homelessness and social isolation conferred little additional risk. Among behavioral determinants, current alcohol users and smokers had higher odds of overdose, whereas former users had decreased odds. Illicit drug use was strongly associated with nonfatal OD for both former and current users. Among clinical variables, tobacco use disorder and alcohol use disorder had strong negative associations with nonfatal OD. We hypothesize that the majority of the patients diagnosed with such disorders were already receiving additional social counseling or clinical support, which helped them build better health and behavioral practices. However, we did not have enough relevant admission data in MIMIC-III to validate this hypothesis; future research is needed to identify the reasons for this observation.

There are several limitations of our study. EHR data are prone to variability by provider documentation and may contain incomplete SBDH information [39]. Additionally, using only ICD-9 codes to identify different medical conditions may lead to inaccurate or misleading values for the corresponding variable. However, structured data often significantly lack SBDH information (only 0.18% for this study), making an NLP-based approach a valuable integration for population studies. Finally, our data had a very low prevalence of nonfatal OD cases (171/48,869, 0.35%), and the MIMIC (ICU) database might not characterize the general outpatient/inpatient hospital setting.

While our study describes an important methodological process that can identify important SBDH factors to consider, which is a necessary first step, further research is needed on subsequent steps on how best to share and translate this information to providers so that they can effectively and actionably use the findings. As our future work, we would like to work on modeling the NLP system predictions for SBDH extraction and how they can be better tied with predictor assessment metrics (eg, OR).

This is the first work to evaluate the risk factors associated with nonfatal OD leading to ICU admissions. Our work concluded that data-driven NLP systems can be largely beneficial in the automatic extraction of SBDH information from unstructured EHR text data. We also showed that analyzing critical care admissions is crucial to better understand the opioid epidemic. Utilizing NLP to leverage the rich EHR notes and more epidemiological studies in critical care settings could be useful for deeper analysis of the OD crisis, leading to the development of better risk assessment tools and effective prevention systems.