Stratified delirium risk using prescription medication data in a state-wide cohort

Abstract

Objective:

Delirium is a common condition associated with increased morbidity and mortality. Medication side effects are a possible source of modifiable delirium risk and provide an opportunity to improve delirium predictive models. This study characterized the risk for delirium diagnosis by applying a previously validated algorithm for calculating central nervous system adverse effect burden arising from a full medication list.

Method:

Using a cohort of hospitalized adult (age 18-65) patients from the Massachusetts All-Payers Claims Database, we calculated medication burden following hospital discharge and characterized risk of new coded delirium diagnosis over the following 90 days. We applied the resulting model to a held-out test cohort.

Results:

The cohort included 62,180 individuals of whom 1.6% (1,019) went on to have a coded delirium diagnosis. In the training cohort (43,527 individuals), the medication burden feature was positively associated with delirium diagnosis (OR=5.75, 95% CI 4.34–7.63) and this association persisted (aOR=1.95; 1.31–2.92) after adjusting for demographics, clinical features, prescribed medications, and anticholinergic risk score. In the test cohort, the trained model produced an area under the curve of 0.80 (0.78–0.82). This performance was similar across subgroups of age and gender.

Conclusion:

Aggregating brain-related medication adverse effects facilitates identification of individuals at high risk of subsequent delirium diagnosis.

1. Introduction

Delirium is an acute confusional state arising in the context of medical illness through an incompletely understood multifactorial process.[1-4] This common and consequential cognitive manifestation of illness is of particular concern to consulting psychiatrists and increasingly discussed in subspeciality geriatrics and critical care literatures,[5-7] as it has been associated with both increased mortality and morbidity.[8] Delirium is associated with longer mechanical ventilation, longer length of hospital stay, longer intensive care unit stays, increased rate of institutional discharge, increased risk of readmission, lower post-hospital health-related quality of life, and post-hospital cognitive deficits.[9-18] Delirium is also associated with increases in medical expense to society,[19-23] is aversive to patients and their families, and is a source of increased caregiver burden for clinical staff.[24-32] Fortunately, multicomponent interventions are able to reduce rates of delirium in some groups.[33-35]

Among the most commonly recognized risk factors for delirium are medication-associated adverse effects.[4,36-39] Medication associated risk is of particular interest because it is modifiable and that modification could occur on the timescale of acute care hospitalization or elective surgery optimization as a portion of recommended prevention programs.[40] Multiple approaches to identifying medications associated with risk have been developed, generally on the basis of expert manual curation.[41,42] In particular, some of these curated lists emphasize anticholinergic medications which are theorized to be of potentially greater relevance to delirium;[43-45] however, studies of human subjects do not consistently identify associations among anticholinergic risk scales and delirium.[46,47]

We have previously described and validated a scalable method for calculating cumulative medication burden for particular categories of adverse effects using FDA labeling and demonstrated that medication burden score could accurately predict risk for future fall.[48,49] Here, we sought to understand whether this approach to medication burden scoring could also predict delirium risk following hospital discharge. We utilized Massachusetts state discharge records to characterize the association automated risk score, in the context of additional sociodemographic and clinical features.

2. Materials and Methods

2.1. Cohort Development

This study of adult patients hospitalized for acute inpatient care drew data from the Massachusetts All-Payer Claims Database (APCD). This database includes claims paid for Massachusetts residents independent of insurance type. The Massachusetts APCD data available for this study did not include Medicare payments and thus the cohort was limited to those less than 65 years of age. All other adults discharged alive from a medical or surgical inpatient admission were considered and all APCD claims data, whether paid through public or private payers, were included. Billing data were used to identify inpatient acute care hospitalizations as the cohort entry defining event. The index visit was defined as the first hospital discharge during the observation period (1 January 2012 and 31 December 2012) with subjects entered the study cohort following their first discharge from acute care. As the study focused on medication associated risk, those who had no medications prescribed at index admission were excluded. The study using non-identifying preexisting clinical data was granted a waiver of informed consent under 45 CFR 46.116.

2.2. Clinical Data Handling

The primary study outcome was delirium diagnosis within 90 days of hospital discharge. Only patients with adequate follow-up to observe the primary outcome were considered. To facilitate interpretation with prior publications, delirium was defined as any of International Classification of Disease (ICD) codes 290.(11,3,41), 293.(0,1,9), or 780.09.[50,51] The original development of this APCD delirium definition considered but did ultimately not include the comparably heterogeneous encephalitis codes.[51,52] Additional demographic and clinical factors were extracted for analysis including age, sex, insurance type (commercial vs. public), the duration of hospitalization (in days), and whether the admission occurred through an emergency department. Coded diagnostic history was used to calculate the Charlson Comorbidity Index and to identify patients with a history of neuropsychiatric diagnosis.[53] These were selected as the intersection of delirium risk factors and available data.[37,46,54-56] Hospital diagnostic claims data were used to identify patients who had a delirium diagnosis during the index hospitalization. Prescription medication data were extracted and handled as described subsequently. For financial analysis all medical expenses incurred after index hospitalization were summed and then normalized to the number of post-hospital days over which these expenses could have been incurred.

2.3. Medication Data and Feature Engineering

Medications in APCD do not distinguish between those prescribed at hospital discharge or following hospital discharge. Therefore, any prescription filled within 15 days of hospital discharge was considered a probable discharge medication and analyzed as a hospital medication for risk scoring. To allow for this 15-day window of medication data to accumulate, patients with events prior to 15 days were excluded from the analysis. Prescribed medications were walked back to prescribed ingredients and those ingredients deduplicated. In other words, a combination medication – for example, one including both acetaminophen and an opioid in a single table – would be expanded to two unique ingredients but multiple prescriptions of a single ingredient – for example, acetaminophen in isolation and in combination – only counted once. Three medication-related features were derived from this unique prescribed ingredient list: (1) the total number of prescribed agents as a simple count, (2) the cumulative Rudolph anticholinergic risk scale score for the list,[44] and (3) a custom delirium medication burden score.[48,49] Although analogous to prior work on falls, this delirium medication burden score assigns different weights to the prescriptible ingredients; as in prior work, we assume additivity of ingredient level burden scores. For example, tiotropium has a calculated delirium burden score of .0005, cyclobenzaprine of 1.78, and alprazolam of 3.38 and thus a patient taking all three would have a crude net burden score of 5.16. The full list of delirium burden scores calculated for each ingredient is shown in Supplemental Table 1.

The delirium medication burden score is a cumulative measure of medication side effects thought by experts to increase the risk of central nervous system (CNS) complications. This measure is the sum of the frequencies of CNS associated side effects reported for each medication a patient is taking. The frequencies of each side effect for each prescribed ingredient were drawn from the SIDER Side Effect Resource databases. The SIDER database maps medications to the frequency of individual side effects associated with those medications using medication labels and post-marketing surveillance data. The custom Burden Score represents an expert-informed summary of how likely a patient is to experience delirium relevant side effects based on the reported frequency of each adverse effect in medication labeling. The minimum score is 0 (no associated adverse effects) with no upper bound, with the assumption that adverse effect frequencies are additive. The assumption of adverse effect additivity is such that if two medications are labeled as being associated with confusion in 10% of patients, a patient treated with both would have a medication burden score of 0.2; if an additional medication also had a 5% frequency of light-headedness, then the burden score would increase to 0.25. This summative medication burden approach has previously been developed and reported for fall risk prediction and independently validated in the APCD.[48,49]

2.4. Statistical Analysis

The full patient cohort was randomly partitioned into a training set for model fitting (70%) and testing set for model prediction evaluation (30%). Descriptive statistics appropriate to each data element type were used to characterize the study cohorts. Multivariable logistic regression was selected as the primary model of the association between medication burden and delirium diagnosis as it provides readily interpretable adjusted odds ratios (aOR) in the training set and extends directly to the prediction of delirium diagnosis in the testing set. Random forest and naive Bayes classifiers were selected as sensitivity checks on the choice of logistic regression as the primary predictive model. The feature variables included in the model were selected as the intersection of variables of known clinical significance and available data elements. As a sensitivity check on the model specification and evaluation of the utility of the medication burden feature of primary interest, forward and backwards stepwise selection by Akaike information criterion was used to evaluate features contributing to model fit.[57] Alternative outcomes of delirium within 60 days of index hospital discharge and 180 days of hospital discharge were evaluated as a sensitivity check on the choice of 90 days as the primary outcome.

Cross-sectional predictive accuracy was assessed by area under the receiver operating characteristic curve (AUC), conventional analysis of sensitivity and specificity, and finally through lift – a comparison between the overall sample event rate and quantiles of predicted risk. Secondary analysis of predictive accuracy across subgroups focused on AUCs by sex and age group. In addition to cross-sectional accuracy, stratified Kaplan-Meier analysis with log-rank testing was used to evaluate the longitudinal risk stratification potential of the resulting model.

In addition to the primary analysis of the multivariable logistic regression model the medication burden score was evaluated isolation. This parallel analysis included calculation of univariable odds of the delirium outcome and AUC of the raw medication burden score as an uncalibrated univariable predictor. Finally, the relationship between the medication burden score and total post-hospital medical expenses per day of potential follow-up available adjusting for private or public insurance, age, and sex was evaluated using multivariable linear regression.

Anticholinergic risk score, total number of medications, medication burden score, Charlson comorbidity index, length of stay, and medical expenses per day were all log transformed due to right skew. All analysis were completed using R v3.

3. Results

The full study cohort included 62,180 individuals with 43,527 randomly allocated to the training set and 18,653 randomly allocated to the test set (Table 1). The full cohort (Table 1, left) was 43.8% male (27,221 of 62,180) and the average age was 51.9 years (SD 8.7) during the index admission. Delirium was present in 0.8% of index admissions (520 of 62,180) admissions and the primary outcome of delirium diagnosis within 90 days of hospital discharge occurred in 1.6% (1,019 of 62,180) of all cases. A history of neuropsychiatric diagnosis in the six months prior to index admission was present in 46.7% (29,016 of 62,180) patients. The average number of medications prescribed during the hospital stay and subsequent fifteen days was 4.50 (SD = 3.18) and the average anticholinergic risk score was 0.38 (SD = 0.97). The average medication burden score was 1.20 (SD = 1.49). Characteristics were similar across the randomly allocated testing and training cohorts (Table 1, center and right).

Table 1:

Sociodemographic characteristics of the full eligible sample (left) contrasted by testing (center) and training (right) cohort.

	Total N = 62180	Test N = 18653	Train N = 43527
Delirium outcome - n (%)	1019 (1.6)	305 (1.6)	714 (1.6)
Private Insurance - n (%)	26274 (42.3)	7890 (42.3)	18384 (42.2)
Patient admitted from ED - n (%)	26647 (42.9)	8037 (43.1)	18610 (42.8)
Prior Neuropsychiatric Diagnosis - n (%)	29016 (46.7)	8712 (46.7)	20304 (46.6)
Delirium at index admission - n (%)	520 (0.8)	171 (0.9)	349 (0.8)
Sex (Male) - n (%)	27221 (43.8)	8138 (43.6)	19083 (43.8)
Age at admission (years) - mean (SD)	51.89 (8.68)	51.91 (8.69)	51.88 (8.68)
Charlson comorbidity index - mean (SD)	4.99 (8.26)	4.95 (8.23)	5.01 (8.28)
Length of stay (days) - mean (SD)	3.51 (2.79)	3.52 (2.78)	3.51 (2.80)
Total prescribed medications - mean (SD)	4.50 (3.18)	4.51 (3.17)	4.50 (3.18)
Anticholinergic risk score - mean (SD)	0.38 (0.97)	0.38 (0.98)	0.38 (0.97)
Medication burden score - mean (SD)	1.20 (1.49)	1.18 (1.48)	1.20 (1.50)

Abbreviations: ED = Emergency Department

The logistic regression model fitted to the training test set is shown in Table 2. The medication burden score was positively and significantly associated (aOR = 1.95, 95% CI 1.31 – 2.92) with subsequent delirium diagnosis after adjusting for clinical and sociodemographic factors, whereas neither medication number nor anticholinergic risk score were significantly associated with delirium. Consistent with expectations based on literature, older age (aOR = 1.02, 95% CI 1.01 – 1.03), male sex (aOR = 1.41, 95% CI 1.21 – 1.64), greater illness burden as captured by Charlson comorbidity index (aOR = 2.53, 95% CI 2.13 – 3.00), history of neuropsychiatric illness (aOR = 1.71, 95% CI 1.43 – 2.06), longer hospitalization (aOR = 2.30, 95% CI 1.66 – 3.18), and admission via the emergency department (aOR=1.50, 95% CI 1.27 – 1.77) were all associated with delirium diagnosis within 90 days (Table 2). Sensitivity analysis of the outcome replacing the 90-day outcome with 60- and 180-day delirium outcomes were consistent with the primary analysis (not shown).

Table 2:

Logistic regression model of 90-day delirium outcome fitted in the training cohort.

Predictors	Odds Ratios	CI	p
Private Insurance	2.9	2.29 – 3.67	<0.001
Patient admitted from ED	1.5	1.27 – 1.77	<0.001
Prior Neuropsychiatric Diagnosis	1.71	1.43 – 2.06	<0.001
Delirium at index admission	2.06	1.35 – 3.15	0.001
Sex - Male	1.41	1.21 – 1.64	<0.001
Age at admission	1.02	1.01 – 1.03	0.001
Charlson comorbidity index ^	2.53	2.13 – 3.00	<0.001
Length of Stay ^	2.3	1.66 – 3.18	<0.001
Total prescribed medications ^	0.84	0.57 – 1.25	0.398
Anticholinergic risk score ^	1.05	0.72 – 1.53	0.812
Medication burden score ^	1.95	1.31 – 2.92	0.001

Abbreviations: ED = Emergency Department

^{^}log transformed

Next, we evaluated the quality of predictions made by the logistic regression model using the independent test set. The test set produced an AUC of 0.80 (95% CI 0.78 - 0.82). The lift in the top decile of predictive risk was 4x, versus .07x in the bottom decile (Figure 1). The highest predicted risk 10% of test set patients accounted for 40% (123 of 305) of cases whereas the bottom 50% of test set patients only included 9% (28 of 305) cases. A sensitivity of 80% occurred at a prediction threshold of 0.012 and produced a specificity of 65%. In secondary subgroup analysis within the test set the AUCs in male patients (0.77, 95% CI 0.74 - 0.80) and female (0.82, 95% CI 0.79 - 0.85) patients were similar as were those across age groups: 0.84 (0.95 CI 0.80 - 0.88) in 35-45-year-olds, 0.79 (0.95 CI 0.75 - 0.83) in 46-55-year-olds, and 0.77 (0.95 CI 0.73 - 0.81) in 56-65-year-olds. For comparison to logistic regression, we trained a random forest and naive Bayes classifier using the same testing cohort and predictive variables and then evaluated these in the training set, producing equivalent AUCs of 0.79 (95 CI 0.77 - 0.81) and 0.79 (95 CI 0.77 - 0.81) respectively. As a final predictive metric, we evaluated the utility of cross-sectional logistic regression predictions in time to delirium outcomes risk stratification. In this analysis, quartile of predicted delirium risk was significantly associated with time to event by log rank test (χ² = 697; p < .001; Figure 2a).

Figure 1:

Predictive lift by decile of predicted risk in the independent testing set

Figure 2:

Kaplan-Meier curves of time to delirium outcome in the testing cohort stratified by (A) quartile of predicted delirium risk and (B) quartile of raw univariate medication burden score.

The medication burden score was independently associated with the delirium outcome (OR=5.75, 95% CI 4.34 – 7.63) in the training set in a univariate model. When the raw burden score was used directly as a delirium risk score it produced an AUC of 0.60 (95 CI 0.56 - 0.64) in the testing set. Quartile of raw medication burden score was significantly associated with time to delirium outcome (χ² = 148; p <.001; Figure 2b). Sensitivity analysis of the multivariable model’s variables by forward and also backwards stepwise selection converged on the same model: inclusion of all variables except total number of medications and anticholinergic risk score—the medication burden score was included as a predictor of the outcome in both automatic selection methods. Finally, in a linear regression of medical cost on medication burden score, adjusted for age, sex, and private insurance, greater medication burden was associated with greater medical costs in both the training (Est = 0.50, 95% CI 0.47 – 0.53) and the test sets (Est = 0.44, 95% CI 0.39 – 0.48). The median daily expenditure in the bottom half of medication burden scores was $70.25 [68.56-71.97], whereas the median daily expenditure in the top 10% of medication burden scores was $178.40 [172.10-185.00].

4. Discussion

In this study of 62,180 individuals discharged from Massachusetts hospitals, a medication burden score based on adverse effect data was significantly associated with the risk of subsequent delirium diagnosis. In a held-out testing set, the predictive accuracy of the model was consistent with those reported in previous delirium prediction studies[58] and prior work demonstrating similar association between medication burden and post hospital fall risk.[48,49] This result supports two potential applications: first, a new approach to medication associated delirium risk captured in the medication burden feature; and second, a potential enhancement to delirium diagnosis risk stratification useful in guiding evidence-based prevention efforts.

Current guidelines and reviews favor multicomponent delirium prevention programs;[35,40,59-62] however, these interventions require ongoing operational resources and in existing reports are not associated with shorter hospital admissions or reduced institutionalization.[34,35] Predictive models of the sort reported here could facilitate more optimal referrals into such efforts by stratifying individuals at greatest risk while the medication burden feature could simultaneously guide a targeted medication risk reduction effort. The AUC of 0.80 in the held-out test cohort is at the upper end of previously reported delirium prediction model AUCs (0.52-0.94) covered in systematic review and that (0.71–0.79) of a more intensive recent machine learning effort.[58,63]

Polypharmacy is a frequently identified risk factor for delirium;[4,36-39] however, crude medication number as a risk factor gives no guidance on prioritizing among medications for targeted risk minimization efforts. In this cohort the association between delirium and the medication burden score was statistically significant whereas the associations between delirium diagnosis and both total number of medications and anticholinergic risk were not. This differential association in favor of medication burden may point the way toward more precise targeted management of medication associated delirium risk. Previous reports have linked specific medication classes (e.g., benzodiazepines) to delirium risk;[64-66] however, class level medication prohibitions are of limited clinical utility as therapeutic intent and adverse effects are correlated at the class level and thus the prescriber’s challenge is often in balancing effect and side-effect when constrained to a single class.[67,68] Among classes, the anticholinergic medications have been of particular focus in the delirium literature and thus have a more developed within-class literature.[44,45] Notably, the lack of association between anticholinergic risk and delirium is consistent with previous studies reporting a lack of association between anticholinergics and delirium.[43,46,47] Whereas prior efforts focused on medication risk stratification required considerable class specific per agent effort, the approach reported here is outcome specific and thereafter scales over all medication without additional effort.[43,44] As comprehensive burden scoring of the sort reported here captures both the additive effect of polypharmacy and the differential contribution of individual agents—both within and across classes—in a fashion which can be scaled across the growing pharmacopeia, burden scoring may be a generally useful approach to an outcome specific quantification of polypharmacy risk and prescription risk mitigation.

In interpreting our results, multiple inherent limitations must be considered. First, we rely on administrative claims of delirium diagnosis which do not accurately identify all clinical case delirium; instead, coded diagnosis typically under report actual delirium.[51,69,70] Undercoding may be particularly true for milder and more subacute presentations.[52] However, we would anticipate that this misclassification would more likely cause us to underestimate the performance of our model due to incorrect classification of true cases as non-cases. Second, it is important to note that the correlation between medication burden score and the delirium outcome reported here does not imply a causal link such that modification of the delirium burden score by modifications to prescribed medication will necessarily alter delirium risk. This correlation could be confounded as modifiable high-risk medications may be proxies for unmodifiable underlying medical comorbidity that itself confers risk (i.e., an example of confounding by indication). Prospective clinical investigation will be required to determine whether a clinical effort to minimize a patient’s burden score actually reduces risk. However, our results do suggest that such studies should be considered as a means of preventing delirium. Ideally future work would include a prospective evaluation of predictive accuracy for an active bedside assessment of the delirium outcome. In parallel, although more complex to calculate, further analysis might consider the role of p450 interactions in delirium medication burdens.[71-73] Finally, the comprehensive state-wide data available for this report do not include elderly patients and thus further research will be needed to evaluate generalizability to the extensive geriatric delirium literature.[5] Risk factors for delirium, and delirium occurrence rates, vary by clinical population such that specific associations among risk factor should not be extended beyond the studied cohort.[51,74] In particular, the present cohort is unlikely to be representative of associations which depend on underlying age-related cognitive change.

4.1. Conclusions

Greater risk for delirium is associated with a greater cumulative medication burden score based on medication side effect data after adjusting for demographics, clinical features, anticholinergic risk score, and total number of medications. This scalable and easily implemented medication burden score could be incorporated into delirium prevention programs and provide an approach to balancing therapeutic intent and delirium risk. Clinical trials are required to establish whether reduction of medication burden scores through targeted pharmacologic optimization efforts translate to a reduction in subsequent cases of delirium.