Abstract
BACKGROUND AND OBJECTIVES: Social risk factors are linked to children’s health, but little is known about how frequently these factors are documented using the International Classification of Diseases (ICD) or whether documentation is associated with health care use outcomes. Using a large administrative database of pediatric hospitalizations, we examined the prevalence of ICD social risk code documentation and hypothesized that social code documentation would be associated with longer length of stay (LOS) and readmission.
METHODS: We analyzed hospitalizations of children ages ≤18 using the 2012 Nationwide Readmissions Database. The following ICD social codes were used as predictors: family member with alcohol and/or drug problem, history of abuse, parental separation, foster care, educational circumstance, housing instability, other economic strain, and legal circumstance. Outcomes included long LOS (top quintile) and readmission within 30 days after discharge. Covariates included individual, hospital, and season variables.
RESULTS: Of 926 073 index hospitalizations, 7432 (0.8%) had International Classification of Diseases, Ninth Revision, social codes. Social code documentation was significantly associated with long LOS. Adjusting for covariates, family alcohol and/or drug problem (odds ratio [OR] 1.65; 95% confidence interval [CI] 1.16–2.35), foster care (OR 2.37, 95% CI 1.53–3.65), other economic strain (OR 2.12, 95% CI 1.38–3.26), and legal circumstances (OR 1.66; 95% CI 1.02–2.71) remained significant predictors of long LOS. Social code documentation was not associated with readmission after adjusting for covariates.
CONCLUSIONS: Social ICD codes are associated with prolonged LOS and readmission in pediatric hospitalizations, but they are infrequently documented. Future work exploring these associations could help to determine if addressing social risk factors in inpatient settings might improve child health outcomes.
Greater social adversity among children is associated with poor overall health,1,2 activity limitations,2 and subclinical biomarkers of disease.3 Consequently, there is increasing interest in identifying and addressing children’s social risks as a core element of pediatric care. The American Academy of Pediatrics recommends screening for social determinants of health during clinical encounters and collaborating with community groups to help address unmet basic needs.4
Despite the growing enthusiasm for collecting this information in clinical settings, few studies have been used to examine whether social risk data are associated with health care use outcomes. This information may be especially relevant to health care systems weighing decisions to invest in social risk screening and intervention programs. In inpatient samples, use research is especially limited, although some recent research reveals that lower socioeconomic status is associated with prolonged hospital stays for those with specific chronic illnesses like asthma5–7 and diabetes.8 In a recent article, researchers also showed that adjusting for social factors improved accuracy of readmission prediction and advocated for including risk adjustment on the basis of these variables.9
Previous research has often relied on insurance status6 as a proxy for socioeconomic status or on area-level socioeconomic indicators such as neighborhood poverty.8 These factors are important predictors of pediatric inpatient outcomes but may not reflect more nuanced individual- or family-level social exposures.
An underexplored data source that could potentially be leveraged to investigate these associations is the International Classification of Diseases (ICD), a coding system used to classify medical diagnoses in health care settings. Although not traditionally used to track social risk factors, the ICD does contain codes that reflect patient and family social situations. Although these codes are underused,10 authors of a recent study found that ICD social code documentation was associated with a higher readmission rate among hospitalized adult patients.11 No previous research of which we are aware has been used to examine the relationship between ICD social risk codes and pediatric use outcomes.
We undertook this study to explore whether social risk data are being incorporated into inpatient pediatric care using existing ICD codes. We also sought to examine relationships between ICD social codes and both length of stay (LOS) and hospital readmissions. We hypothesized that, similar to previous work, ICD social codes are rarely documented among pediatric hospitalizations10 but, when documented, are associated with an increased likelihood of long LOS and readmission.
Methods
We analyzed data from the 2012 Nationwide Readmissions Database (NRD) component of the Healthcare Cost and Utilization Project, version 2. The Agency for Healthcare Research and Quality oversees the NRD, and in 2012, it included data from 18 states on hospital readmissions in both children and adults. The NRD, for which the unit of analysis is a hospitalization (as opposed to a patient), contains both individual-level variables, including demographic, diagnosis, LOS, and mortality data, as well as hospital-level variables. It is weighted to enable nationally representative analyses. Additional information on the NRD data collection methods is available elsewhere.12
Sample
In this analysis, we included hospitalizations of patients ages ≤18 during their index admission. We excluded encounters in which patients either died, were transferred in or out of the hospital (or when it was unclear whether a transfer in or out occurred), left against medical advice, were discharged to an unknown destination, or had a same-day event wherein the discharge date for 1 inpatient stay was the same as the admission date for another stay (Fig 1A).
Sample selection for LOS analyses. A, Sample selection of primary sample for LOS analysis. B, Sample selection of sensitivity analysis sample for LOS analysis.
In a sensitivity analysis, we included hospitalizations of patients readmitted nonelectively (as defined by an NRD-provided elective admission indicator variable) after the index admission using the same exclusion criteria (Fig 1B).
When analyzing 30-day readmissions, we excluded index hospitalizations with a December discharge date because readmissions within 30 days that occurred in January of the following year could not be tracked. See Supplemental Table 4 for details on the NRD variables used.
Variables
Predictors
On the basis of previous literature that has linked pediatric health outcomes with specific social risks,13–18 we chose the following social risk factors and corresponding International Classification of Diseases, Ninth Revision (ICD-9), codes as independent variables: family member with alcohol and/or drug problem, history of being a victim of abuse, parental separation (divorce or estrangement), foster care (including disruption due to child welfare custody), educational circumstances, housing instability, other economic strain, and legal circumstances. Please see Supplemental Table 5 for details of the ICD social codes.
These ICD social codes were derived from each pediatric hospitalization record. We separately evaluated associations between ICD-9 social code documentation for each of these 8 domains with long LOS. We additionally evaluated associations with a binary indicator of whether any 1 of these social codes was assigned during the index hospitalization.
Outcomes
The distribution of LOS is not normal, so as done in previous studies,7,8,19 we used percentiles, specifically quintiles, to define long LOS. Outcomes included the hospitalization being in the top quintile for hospital LOS (>3 days) as well as readmission within 30 days of the index admission.
Covariates
Covariates were selected on the basis of their expected association with pediatric readmission and LOS.6,7,20–25 These included patient age, sex, zip code income quartile, insurance payer, season, number of chronic medical conditions, size of hospital, type of hospital (eg, private nonprofit), teaching hospital status and/or location (eg, urban nonteaching), severity of symptoms based on All Patient Refined Diagnosis Related Group (APR-DRG) codes, and likelihood of mortality on admission based on APR-DRG codes.
Given the varying prevalence of ICD social codes in hospitalizations depending on the reason for hospitalization,10 we also created an indicator variable for each NRD assigned major diagnostic category (MDC) (eg, “diseases and disorders of the respiratory system,” “diseases and disorders of the circulatory system”26). We used the originally assigned MDC category when there were ≥50 patients in a category and collapsed the less densely populated MDC categories into an “all other” category for both sample size and anonymity reasons.
Some hospitals may have a higher prevalence of social code use compared to others because of internal hospital practices or incentives. To account for this, we developed a hospital social code adjustment weight to control for the overall frequency of social code documentation in hospitals. This variable was constructed by dividing the weighted number of admissions of each hospital by the weighted number of admissions for which a social code was documented at that hospital.
Please see Supplemental Table 4 for details on how all variables from the NRD were derived and, when relevant, recoded.
Data Analysis
Weighted distributions of covariates among hospitalizations with and without social codes and weighted distributions of the presence of any social code by covariates were calculated and compared by using χ2 testing. The raw and weighted prevalence outcomes of each individual social code in the sample were also calculated.
Logistic regression was used to estimate the odds of having a long LOS on the basis of the presence of ICD social codes in the index admission. Model 1 was a univariate model; model 2 added median neighborhood income and insurance status, the measures commonly used to denote social risk factors in previous studies; and model 3 added all other covariates except for the hospital social code adjustment weight, which was added in the final model (model 4).
This stepwise regression method was chosen a priori to help investigate potential pathways through which social risk factors may operate. If a social risk factor was found to be a significant predictor of long LOS when adjusting for neighborhood income and insurance status, but not when adjusting for health status variables, it could suggest that that factor’s association with long LOS is confounded with and accounted for by other established predictors of LOS, such as severity of symptoms. Alternatively, if a social risk factor was found to be significant when controlling for all covariates, it might represent a unique risk factor above and beyond traditional predictors. In essence, this type of modeling helps to investigate potential mechanisms through which certain predictor variables may be influencing others; although, because the data are from an administrative database, causality cannot be established.
The longer a patient stays in the hospital, the more opportunities there are for a team member to document social factors, which could lead to reverse causation. To investigate this, we conducted a sensitivity analysis among those patients with a nonelective readmission using the social codes documented in patients’ index admission to predict long LOS (highest quintile of LOS was >5 days in the sensitivity sample, technically 21%) in the subsequent, nonelective admission. This ensured that for the second admission, the social factor was documented before admission. Because of sample size restrictions, we only used the binary indicator of having any social code as a predictor.
Cox proportional hazard regression was used to estimate the risk of 30-day nonelective readmission by using only the aggregate binary social code indicator because sample sizes for analyzing individual social codes with readmission as an outcome were small. We used the same covariates in each model as for the logistic regression analyses.
All models were nationally weighted by using NRD-provided sampling weights. All analyses were performed by using Stata version 15.1 (Stata Corp, College Station, TX). This study was deemed exempt by the University of California, San Francisco, Institutional Review Board.
Results
Of the 1 091 255 hospitalizations of children ages ≤18, a total of 967 596 were index admissions. Of those, 926 073 (95.7%) met inclusion criteria (Fig 1). Table 1 and Supplemental Table 6 indicate the prevalence and odds ratios (ORs) of having social codes by covariates. Social codes were significantly more likely to be documented in hospitalizations among those living in lower median-income quartile areas and less likely to be documented among hospitalizations of patients with private insurance. They were also more commonly documented among older children, girls, and for those with chronic diseases and less frequently documented in smaller hospitals. Hospitalizations with minor-severity APR-DRG codes were less likely to have a social code than those with moderate or major severity. Additionally, hospitalizations with minor likelihood of mortality (according to APR-DRG codes) were more likely to have social codes than those with extreme likelihood of mortality. Social code documentation also differed significantly by discharge quarter and MDC category. There were no significant differences in the presence of social code documentation based on hospital ownership or hospital location and/or teaching status.
Sample Characteristics by Presence of ICD Social Code (N = 926 073)
The frequency of social code documentation is displayed in Table 2. In our analytic sample, 7432 (0.80%) had at least 1 ICD social code documented (all other results mentioned represent weighted data). The prevalence of specific types of social codes ranged from a low of 0.02% for parent separation to 0.28% for victims of abuse.
Number and Percentage of Hospitalizations With Social Code
Every individual social code and the aggregate measure of social code documentation were predictive of long LOS even when controlling for neighborhood income and insurance status (Table 3). When controlling for all covariates, the presence of any social code was associated with a 1.40-times (95% confidence interval [CI] 1.12–1.74) increased odds of having a long LOS. Additionally, when controlling for all covariates, family alcohol and/or drug problem (OR 1.65; 95% CI 1.16–2.35), foster care (OR 2.37; 95% CI 1.53–3.65), other economic strain (OR 2.12; 95% CI 1.38–3.26), and legal circumstances (OR 1.66; 95% CI 1.02–2.71) all remained significantly associated with long LOS.
ORs of Having Health Care Outcome by Presence of Social Code
There were 851 640 hospitalizations eligible for readmission analysis after excluding December index admissions. In Cox proportional hazards models, having a social code was associated with an increased risk of 30-day readmission even when controlling for neighborhood income and insurance status (hazard ratio 2.67; 95% CI 2.25–3.16; Table 3). However, when additionally controlling for patient, hospital, and season characteristics (model 3), the effect became insignificant.
Our sensitivity analysis sample consisted of 57 586 children (Fig 1B). The social code binary indicator from the index admission was related to long LOS during subsequent readmission when controlling for neighborhood income and insurance status (OR 3.32; 95% CI 2.79–3.97; Table 3). However, this association disappeared when controlling for other covariates.
Discussion
Using data from a large sample of pediatric hospitalizations, we found that ICD social risk codes are documented infrequently but, when present, are associated with long LOS. These associations remained even after controlling for a set of factors already known to influence LOS, including measures of insurance status and neighborhood income. Having a social code documented was also predictive of earlier readmission within the 30-day period after the index hospitalization when controlling for insurance status and income quartile. This is the first study of which we are aware that links these ICD social risk codes to any pediatric hospitalization outcomes.
Our finding of a low prevalence of ICD social code documentation aligns with previous work.10 Health services researchers have shown that ICD codes in large administrative data sets like the NRD have high specificity but that sensitivity ranges on the basis of the diagnosis.27–29 Also similar to earlier work, specific populations were more likely to have social codes recorded.10 For example, social code documentation was common among admissions for mental health conditions. This is consistent with findings that children with certain social risk factors have a higher likelihood of experiencing mental health problems during childhood.30,31 Alternatively, this could represent providers conducting a more thorough exploration of social stressors among patients with mental health conditions or a higher likelihood of mental health professionals using these social codes.
Despite the overall low prevalence of ICD social codes, we found that they are associated with long LOS. As the influence of social adversity on health is well documented in previous work,32–35 these social codes may therefore be appropriately capturing patients with greater social disadvantage who are experiencing higher risks of disease and barriers to care. They may subsequently be presenting with severe symptoms requiring longer recovery periods. An alternative explanation is that social circumstances delay discharges. For instance, if a child needs foster placement, it may take more time to discharge the child from the hospital. In either case, we found social codes were associated with health care use. Identifying social risk factors early during hospitalizations and targeting them for interventions may represent an opportunity to decrease use in certain circumstances.
On the basis of a growing awareness that social risk factors have an immediate relevance to discharge planning and may present opportunities to reduce future use, new research might leverage opportunities like the Centers for Medicare and Medicaid Services’ Accountable Health Communities and Comprehensive Primary Care Plus demonstration programs to incentivize social risk documentation. These opportunities are augmented by updates available in the International Classification of Diseases, 10th Revision, that provides more specific social risk codes. For example, we were unable to analyze food insecurity using ICD-9 codes because the code that specifies effects of hunger (994.2) was too broad, including diagnoses of famine and conditions related to starvation. The International Classification of Diseases, 10th Revision, alternatively, has a code for “lack of adequate food and safe drinking water” (Z59.4).
Screening in inpatient settings may also circumvent existing barriers found in the outpatient setting. For example, although staff working in ambulatory settings may be concerned about time constraints,36 there is typically more time for collecting social data and more staff available for related interventions in the inpatient setting. As health systems consider documentation strategies, it will be important to ensure that documentation practices involve patient input and minimize possible unintended consequences as described by Gottlieb and Alderwick.37
Our study has several limitations. First, in an attempt to avoid overinflating associations between social codes and health care use outcomes driven by a subset of patients with multiple hospitalizations, we only analyzed index admissions in our main analyses. This may have limited associations that are potentially seen when analyzing multiple readmissions. Additionally, pediatric readmissions are relatively rare, so associations with even plausible risk factors are subject to the statistical challenges of studying rare outcomes.
The small number of hospitalizations with social code documentation highlights that our findings are suggestive but not definitive. The low overall social code use in the NRD also likely affects the generalizability of our results by underestimating both the prevalence of social risk factors in the pediatric population and the extent to which social risk factors are discussed in the context of health care delivery. For example, although previous research has suggested that roughly 12.3% of children live with parents with a substance use disorder,38 our study documented this in only 0.11% of hospitalizations. It is conceivable that social codes are only being used for the most severe cases of social disadvantage. Alternatively, social factors may be documented more often in patient notes, but not via ICD codes, as a recent study found.11 Because ICD social codes are much more widely accessible than other elements of the electronic health record, efforts to incentivize social code use could powerfully improve the availability of social risk data for both clinical and research uses.
Our study does not enable us to attribute causal effects. In fact, our sensitivity analysis suggests reverse causation might account for some of the associations we found because the overall association between presence of a social code and LOS was weakened in these models. Reverse causation could also affect more upstream causes of hospitalization. For instance, the expense of caring for sicker children with more chronic medical conditions could deplete families’ financial resources, subsequently leading to worse illness (eg, through fewer resources to pay for medications) and longer LOS.
Conclusions
There is growing interest in capturing social information in health care settings.39 ICD social risk codes offer an opportunity to incorporate this critical data on children’s social background and environment into health records by using an existing, standardized medical vocabulary that is ubiquitous across the US health care system and that could maximize the utility of collected data to clinical providers. Yet there remains an important gap between the known prevalence of these social risk factors and their clinical documentation. Our finding that social risk factors as documented using ICD codes are associated with health care use outcomes should prompt health care practitioners and payers to consider more systematic approaches to social risk data collection and interventions. Documenting social risk factors consistently could be a useful tool for health systems to help identify families for potential interventions to reduce health care use and improve health.
Footnotes
FINANCIAL DISCLOSURE: The authors have indicated they have no financial relationships relevant to this article to disclose.
FUNDING: Dr Pantell receives support from the National Institutes of Health Loan Repayment Program (award 1 L60 MD013257-01).
POTENTIAL CONFLICT OF INTEREST: The authors have indicated they have no potential conflicts of interest to disclose.
References
- Copyright © 2020 by the American Academy of Pediatrics
In this issue
