Applying machine learning to understand factors predicting pharmacotherapy for mental health support among adults with intellectual and developmental disabilities

Abstract

Americans with intellectual and developmental disabilities (IDD) experience health disparities, including in their mental health. This often leads to disproportionate use of psychotropic medications, sometimes leading to serious side effects. We used machine learning to analyze an integrated dataset (years 2018–2022) from one U.S. state with 2907 observations and 850 variables to determine what factors were most predictive of pharmacotherapy use to support mental health needs among people with IDD. Our algorithm performed strongly, with the presence of mood, anxiety, and psychotic disorders, documented behavioral support needs, and overall support needs all contributing strongly to the algorithm’s accuracy. Implications for social workers and other mental health professionals are discussed.

Background

As of 2019, about 2.2% of the U.S. adults population was identified as having an intellectual and/or developmental disability (IDD), comprising roughly 7.39 million people (Larson et al., 2022). Among Americans aged 3–17, that prevalence was higher, at about 8.56% as of 2021, and that number was a significant increase from 7.40% just three years prior (Centers for Disease Control and Prevention, 2023). Thus, Americans with IDD represent a substantial and potentially growing segment of the U.S. population.

Intellectual and developmental disabilities are two distinct, yet related conditions. Intellectual disabilities manifest prior to age 22 and are characterized by significant impacts to both intellectual functioning and functional behavior (American Association on Intellectual and Developmental Disabilities, n.d.). Developmental disabilities capture a wider array of conditions that encompass intellectual, physical, language, or behavioral limitations that affect a person’s daily functioning (Centers for Disease Control and Prevention, 2022). ID and DD are both lifelong conditions. Though their definitions are distinct, in research and practice it is commonplace for ID and DD to be considered together (as IDD).

Mental health among people with IDD

In 2023, the National Institutes of Health (NIH) declared people with disabilities (including, but not limited to, IDD) as a population that experiences health disparities, thereby opening up new opportunities to grow research and answering the calls of advocates who had been making this argument for many years (NIH, 2023). Mental health disparities have long been documented as one area in which health disparities are particularly stark, specifically for people with IDD (Deb et al., 2007). For instance, Emerson and Hatton (2007) noted a 40% prevalence rate for mental health conditions among adults with IDD, and Esler and colleagues found prevalence to be 44.8% (2019). By comparison, the National Alliance on Mental Illness estimates that about 21% of U.S. adults live with a mental health condition (National Alliance on Mental Illness [NAMI], 2022. As in the general population, the most common mental health concerns documented for people with IDD are mood disorders (particularly major depression), and anxiety disorders (Hassiotis & Ouellette-Kuntz, 2014).

Though high rates of mental health conditions among adults with IDD have been noted for years, there is also evidence that adults with IDD may not be receiving the support they need to manage their mental health. Both people with IDD and mental health providers have expressed difficulties when adults with IDD seek care, including inadequately prepared practitioners, poor service quality, and limited understanding of IDD (Whittle et al., 2017), findings that have been replicated among autistic adults, a subset of the IDD population (Maddox et al., 2019). Perhaps due, in part, to these difficulties in accessing disability-competent primary mental health care, adults with IDD often rely on hospital emergency departments when mental health concerns arise (Lauer et al., 2019) and may make repeat visits to emergency departments for concerns that could otherwise be handled in community-based primary care settings (Lunsky & Balogh, 2010). People with IDD are also more likely than members of the general population to experience inpatient psychiatric treatment (Lauer et al., 2019).

Recent literature has identified a number of possible explanations for high prevalence of mental health conditions among adults with IDD, including the potential contribution of social exclusion, stigma, and discrimination which can increase vulnerability to mental health concerns (Harris & McMahon, 2018). Emerson (2016) noted that cognitive impairments that are often present among people with IDD may make it more challenging to process social exclusion, and thereby contribute to the presentation of mental health conditions. Finally, Deb et al. (2007) are among the authors who have pointed to a lack of professionals trained to recognize and support the mental health challenges of people with disabilities, whose mental health struggles are often dismissed by professionals as simply being a part of one’s disability.

Pharmacotherapies to support mental health among adults with IDD

Considering the high prevalence rates, and reported difficulties with accessing and utilizing disability-competent mental health services, it is of little surprise that use of pharmacotherapies to treat mental health conditions among people with IDD is common. Though rates vary by study and country, some recent studies have found that about 40% (Kapp et al., 2014) to as high as 50% (Iacono & McCarthy, 2020) of adults with IDD use some type of psychotropic medication to manage mental health concerns. This compares to an estimated 20% of the overall adult U.S. population that has used medication to manage mental health in the past year (NIMH, 2023).

Though pharmacological interventions are often considered as a part of an individualized treatment plan for managing one’s mental health, the disproportionate use of pharmacotherapies among people with IDD is of concern due to the potential for long term side effects (Hollins & Attard, 2012). Other researchers have noted that polypharmacology is fairly common among people with IDD, particularly in the simultaneous use of antipsychotics and antidepressants, often without proper monitoring of dosages and side effects (Sullivan et al., 2017). Such observations have led authors to call for increased implementation of individualized treatment planning (Einfeld et al., 2016) and increased use of interdisciplinary approaches in the treatment of mental health among people with IDD (McCarthy et al., 2019).

Along with the risks associated with long- and short-term use of pharmacotherapies, there are some potential benefits. Pharmacotherapies can provide relatively rapid symptom improvement, which can be quite helpful, especially in instances of mental health conditions are accompanied by escalated behavioral challenges. A recent meta-analysis of psychologically-based therapies among people with IDD found that, while therapies did have a small but significant effect overall (Tapp et al., 2023), responsiveness to therapies often took longer than expected for people with IDD, meaning that pharmacotherapy use may remain preferred for early treatment, while the effects of talk therapies take hold.

The role of social workers in supporting mental health of people with IDD

As the largest professional provider of mental health services and supports in the United States (U.S. Health Resources and Services Administration, 2023), social workers are in an ideal position to support the mental health of people with IDD. Training in the provision of mental health services is a keystone of most Master’s-level social work education programs (Bland et al., 2021), and the top area of specialization among Master’s-level social workers, nationally in the U.S (Council on Social Work Education, 2023). In addition to providing a large portion of the nation’s mental health workforce, social workers are trained to intervene in structural injustices at community and policy levels, grounded by their ethical obligation to pursue social justice (National Association of Social Workers, 2021). This training situates them ideally to pursue solutions that will not only help to support the mental health of people with IDD individually, but also on structural levels.

Despite their large presence in the mental health workforce and multi-faceted training, social workers’ presence as providers of mental health services for people with IDD remains somewhat limited, as has been noted by both practitioners and advocates (e.g. Coyle, 2025; Edelman, 2024). Part of the reason for this is likely the dearth of opportunity to learn about and engage with people with disabilities in the scope of a student’s typical social work education (Kim & Sellmaier, 2020; Ogden et al., 2017). Given this, it is little wonder that so few social workers pursue a career working with people with disabilities, which leaves a service gap that is often filled by other professionals, including primary care physicians who often serve as the main gateway to mental health services (Olfson, 2016), particularly in absence of access to qualified mental health service providers to refer to (Cunningham, 2009). This reliance on medical professionals as providers of mental health services naturally results in heavier than expected use of pharmacotherapies. As such, it is important that we understand the drivers of pharmacotherapy reliance for people with IDD so that we can work toward better-training the social work workforce to intervene with people with IDD in ways that will enable balance between socially driven and more medicalized approaches to treatment.

Machine learning overview

A key feature of this article is the use of machine learning as one of the core methodological drivers. Though machine learning is an established and rapidly expanding approach to data analysis, its use in the disability and mental health fields has been limited thus far. Therefore, we are offering this brief summary of the method to help ground readers in a basic understanding, though it bears mentioning that machine learning is a highly complex and nuanced topic that cannot be adequately described in a short introduction.

Originally derived as a branch of computer science, machine learning is an inductive approach, meaning it is driven by different logic and assumptions than traditional parametric statistical methods that dominate the social sciences, which are deductive in nature. In traditional deductive methods, a researcher will develop a hypothesis to be tested with outcome and predictor variables that are hand-picked by the researcher, a priori, inherently carrying the bias of prior findings, theories, or values. Conversely, a machine learning analyst will begin with a large dataset, containing as many observations and as many variables as possible to predict an outcome. Through an inductive process of testing random subsets of the dataset thousands of times, the dataset is used to inductively construct an algorithm that explains the chosen outcome variable, including an understanding of which variables contribute the most to the accurate prediction of the outcome (Breiman, 2001). Thus, the data, not a researcher’s a priori choices, drives the analytic process, and ultimately the findings.

Though still relatively novel in many sub-fields of the social sciences, the use of machine learning has become increasingly common in recent years. Among the forerunners were criminal justice, where machine learning was adopted relatively early to predict recidivism (Beck, 2019), and education, where machine learning has regularly been used to monitor student progress (Baker & Inventado, 2014; Chung & Lee, 2019). The uptake of machine learning to understand mental health outcomes has been steady, with recent systematic reviews summarizing the state of studies using the methodology to date (Shatte et al., 2019; Thieme et al., 2020). The use of machine learning applications in the IDD field has been less frequent, though it is growing (Santiago & Smith, 2019). Notable machine learning studies in the IDD field include investigations of predictors of employment or other day activity participation (Broda et al., 2021), patterns of health problems (Bishop-Fitzpatrick et al., 2018), autism detection (Thabtah & Peebles, 2020), and patterns of COVID-19 infection (Broda et al., 2024).

Study justification and aims

As the United States’ largest professional bloc providing behavioral health services (U.S. Health Resources and Services Administration, 2023), social workers are in a unique position to help improve mental health outcomes for adults with IDD. However, recent research has suggested that there are difficulties in interdisciplinary teaming (McCarthy et al., 2019) and the development of individualized treatment plans (Einfeld et al., 2016). These challenges may inhibit access to, and effectiveness of mental health services for adults with IDD. They likely contribute to the heavy reliance on pharmacotherapies to treat mental health conditions among people with IDD, often with improper monitoring.

To this end, this paper seeks to build a greater understanding of the factors that contribute to pharmacotherapy use among adults with IDD. If we can better understand the factors driving pharmacological intervention, then social workers and other mental health professionals may be better positioned to design interventions that can assist people with IDD to live good lives in their communities, while reducing the risks posed by long-term medication use. Thus, our study aims are twofold:

1. To test the feasibility of a machine learning algorithm to predict pharmacotherapy for mental health conditions among a large random sample of adults with IDD who use state-funded services.

2. To identify the factors that most accurately contribute to the understanding of medication use for treatment of mental health conditions among a large random sample of adults with IDD.

Methods

Ethics statement

This research was conducted after review and approval by the Institutional Review Board (IRB) of the authors’ university. Ongoing review is conducted as required. The authors have no conflicts to declare.

Data sources & key variables

This study used three major datasets from one U.S. state, which were merged together at the individual service user level and combined over five years (encompassing cohorts from 2018–2019 through 2022–2023) to create one dataset that was used for this study. Each of the datasets and notes on the integration process are found in this section.

National core indicators-intellectual and developmental disabilities (NCI-IDD)

The NCI-IDD (National Core Indicators, 2025) is a voluntary collaboration between participating states, and the creators of the survey: Human Services Research Institute and the National Association of Directors of Developmental Disabilities Services (HSRI & NASDDDS). A nationally validated instrument, the NCI-IDD is administered as a face-to-face survey (either in person or via videoconference) between a trained interviewer and an adult with IDD who used state-funded services and supports in a particular year. Random, independent samples are drawn each year from the pool of adults who use state-funded Long Term Services and Supports. In [state name redacted], about 800 randomly selected adults with IDD participate in the NCI-IDD each year.

There are three sections in the NCI-IDD. First, a background section is completed by a case manager or other person who knows the person with IDD well and has access to their file data. The background section provides demographic information, information about health and mental health diagnoses, and information about the types of services and supports a person has used during the survey year. Second, Section I of the in-person interview must be answered by the person with IDD directly (no proxy response permitted) and contains subjective questions about one’s life satisfaction, satisfaction with services, and related subjective outcomes. Finally, Section II of the in-person interview can be answered by the person with IDD or a proxy (often a parent or a direct support professional), and contains objective information about choice making, social relationships, and community participation. Variables from all three sections of the NCI-IDD were used in our analyses.

Of particular note, the background section of the NCI-IDD includes several mental health-related variables, including the outcome variable for this study. Specifically, the item that we used as our outcome variable asked whether the participant with IDD “Takes medication for mood, anxiety, and/or psychotic disorders?” Response options were Yes, No, or Don’t Know. For our analyses, the Don’t Know responses were treated as missing data. The respondent to this item was typically a case manager or other person with direct knowledge of the participant with IDD, and was completed ahead of the face-to-face interview, usually using the person’s records. All other NCI-IDD variables were included in the dataset for analyses to train and test our algorithm.

Supports intensity scale - adult version (SIS-A)

The SIS-A is a standardized, nationally validated measure (Thompson et al., 2015) designed to assess the support that a person with IDD needs to live well in the community. In [state name redacted], the SIS-A is used as one element that may help determine the supports that a person with IDD is eligible to receive, though it is not determinative. Other states may use the SIS-A differently (such as in a way that is determinative is setting budgets for support needs), though the administration of the SIS-A would not change since it is a standardized instrument. It is used in 33 states and two Canadian provinces, though we only used data from one state in our study.

The SIS-A is administered face to face, most typically by a trained case manager and the person with IDD, and sometimes with their caregiver. The administration time varies by individual, depending on communication method and other factors. Questions cover topics such as activities of daily living, personal advocacy, and exceptional medical and behavioral support needs.

We used all the individual variables from the SIS-A in our study. In addition, we used three constructed variables: 1) the Support Needs Index (a composite score indicating total support needs), 2) Exceptional Medical Support (an index used in our state to identify people with high medical support needs), and 3) Exceptional Behavioral Support (an index used in our state to identify people with high support needs for challenging behavior).

Medicaid LTSS expenditures

We used a single variable, representing the state’s total Medical LTSS expenditure for a particular individual in the particular year of interest. These data were provided to us by the state’s Medicaid agency.

Accounting for missing data

As the data used here were part of a larger secondary survey, missing data was present for many variables. Table 1 includes a category called “Missing” that identifies the number of observations missing for each variable. Overall, levels of missingness were quite low, below 5% for all but two variables included in the table: age, which had 25% of observations marked as missing, and level of ID, which had a missing rate of 9%. Details on the analytic approach to handling missing data can be found in the analysis section below.

Table 1.

Descriptive Statistics for NCI-IPS Sample, by Mental Health Medication Status.

	No (N = 1220)	Yes (N = 1687)	SMD
Demographic Variables
Age at end of survey year			0.081
Mean (SD)	41.0 (15.5)	42.3 (15.9)
Missing	330 (27.0%)	400 (23.7%)
Race/ethnicity: White			0.092
No	492 (40.3%)	606 (35.9%)
Yes	725 (59.4%)	1079 (64.0%)
Missing	3 (0.2%)	2 (0.1%)
Race/ethnicity: Black or African-American			0.086
No	800 (65.6%)	1175 (69.7%)
Yes	417 (34.2%)	510 (30.2%)
Missing	3 (0.2%)	2 (0.1%)
Race/ethnicity: Hispanic/Latino (Mexican, Mexican-American, Chicano, Puerto Rican, Cuban, or Other Spanish/Hispanic/Latino)			0.04
No	1197 (98.1%)	1648 (97.7%)
Yes	20 (1.6%)	37 (2.2%)
Missing	3 (0.2%)	2 (0.1%)
Gender			0.069
Male	712 (58.4%)	1021 (60.5%)
Female	508 (41.6%)	659 (39.1%)
Other	0 (0%)	2 (0.1%)
Missing	0 (0%)	5 (0.3%)
Level of Intellectual Disability
Overall SMD			0.395
Mild	261 (21.4%)	550 (32.6%)
Moderate	399 (32.7%)	619 (36.7%)
Severe	215 (17.6%)	225 (13.3%)
Profound	160 (13.1%)	93 (5.5%)
Unspecified	53 (4.3%)	51 (3.0%)
Diagnosis Unknown	6 (0.5%)	1 (0.1%)
Missing	126 (10.3%)	148 (8.8%)
SIS-A Composite Variables
Support Needs Index			0.037
Mean (SD)	102 (10.1)	103 (9.77)
Extraordinary Behavioral Support Score			0.084
Mean (SD)	2.33 (3.50)	2.63 (3.81)
Extraordinary Medical Support Score			0.031
Mean (SD)	3.35 (3.71)	3.24 (3.49)
Important Distinguishing Variables
Mood Disorder			1.229
No	1132 (92.8%)	793 (47.0%)
Yes	54 (4.4%)	846 (50.1%)
Missing	34 (2.8%)	48 (2.8%)
Anxiety Disorder			0.911
No	1115 (91.4%)	952 (56.4%)
Yes	74 (6.1%)	678 (40.2%)
Missing	31 (2.5%)	57 (3.4%)
Person takes medication for behavioral challenges			0.897
No	1079 (88.4%)	844 (50.0%)
Yes	134 (11.0%)	794 (47.1%)
Missing	7 (0.6%)	49 (2.9%)
Psychotic Disorder			0.692
No	1178 (96.6%)	1268 (75.2%)
Yes	16 (1.3%)	369 (21.9%)
Missing	26 (2.1%)	50 (3.0%)
Disruptive behavior			0.77
No support needed	881 (72.2%)	615 (36.5%)
Some support needed;	255 (20.9%)	735 (43.6%)
Extensive support needed;	77 (6.3%)	308 (18.3%)
Missing	7 (0.6%)	29 (1.7%)
Behavior that is destructive or harmful to others			0.625
No support needed	989 (81.1%)	897 (53.2%)
Some support needed;	168 (13.8%)	536 (31.8%)
Extensive support needed;	51 (4.2%)	222 (13.2%)
Missing	12 (1.0%)	32 (1.9%)
Behavior Disorder			0.613
No	983 (80.6%)	901 (53.4%)
Yes	217 (17.8%)	748 (44.3%)
Missing	20 (1.6%)	38 (2.3%)
No other disabilities other than ID			0.509
No	933 (76.5%)	1494 (88.6%)
Yes	179 (14.7%)	30 (1.8%)
Don’t Know	3 (0.2%)	2 (0.1%)
Type of Residence			0.494
ICF/IID, 16 or more residents with disabilities	0 (0%)	1 (0.1%)
Group living setting, 2–3 people with disabilities	61 (5.0%)	122 (7.2%)
Group living setting, 4–6 people with disabilities	277 (22.7%)	626 (37.1%)
Group living setting, 7–15 people with disabilities	55 (4.5%)	81 (4.8%)
Lives in own home or apartment; may be owned or rented, or may be sharing with roommate(s)or spouse	86 (7.0%)	146 (8.7%)
Parent/relative’s home (may include paid services to family for residential supports)	587 (48.1%)	437 (25.9%)
Foster care or host home (round-the-clock services provided in a single-family residence where 2 or more people with a disability live with a person or family who furnishes services)	53 (4.3%)	104 (6.2%)
Foster care or host home (round-the-clock services provided in a single-family residence where only one person with a disability lives with a person or family who furnishes services – sometimes called shared living.)	95 (7.8%)	157 (9.3%)
Homeless or crisis bed placement	0 (0%)	2 (0.1%)
Other	1 (0.1%)	3 (0.2%)
Don’t Know	1 (0.1%)	0 (0%)
Missing	4 (0.3%)	8 (0.5%)
Support Needs Index			0.037
Mean (SD)	102 (10.1)	103 (9.77)

SMD = standardized mean difference, calculated between Beh. Meds (yes) and Beh. Meds (no) groups. NCI-IPS = National Core Indicators – In Person Survey.

Data management and integration

A key methodological feature of this study was the integration of the three datasets noted above over multiple years. We used a unique identifier that was present on all three of the data sources as our key for the data integration process. Using that identifier, we were able to integrate the NCI-IDD, SIS-A, and Medicaid LTSS data for a particular individual in a particular year. Since all people who use LTSS in [state name redacted] have Medicaid expenditure and SIS data on file, our main constraint was the size of each year’s NCI-IDD sample, which is generally about 800 people per year. Using the NCI-IDD as a base, we were able to merge SIS-A and Medicaid LTSS expenditures for all participants. Next, we merged data from five annual cohorts (2018–2019 through 2022–2023) to arrive at our final dataset. In total, our data have 850 variables and a total of 2907 observations.

Analysis approach

Feature engineering

To conduct feature engineering and data processing, we used the tidymodels package (Kuhn & Wickham, 2020) in R (R Core Team, 2020). We scaled all continuous variables in the NCI-IDD, SIS, and Medicaid expenditures to have a mean value of 0 and a standard deviation of 1. For categorical variables, we created a unique dummy variable for each category. Missing values were replaced using median imputation for all continuous variables, and k-nearest neighbors imputation for all categorical variables (M. Kuhn & Johnson, 2013).

Analytic process

After applying the feature engineering and data processing features outlined above, we conducted our analyses in three steps: 1) classification trees using the “rpart” package in R (Therneau & Atkinson, 2023), 2) random forest models, via the “ranger” package (Wright & Ziegler, 2017), and 3) boosted regression trees using the “xgboost” package (Chen et al., 2023). This multi-model approach allows us to compare the performance across different models. The three models chosen are well established and widely used in predictive modeling (e.g. M. Kuhn & Johnson, 2013). For this study, we defined model performance as the overall classification accuracy of a model under the Receiver Operating Characteristic Area Under the Curve (ROC AUC). Since it included the aggregation of thousands of randomly constructed classification trees, we anticipated that the random forest approach would achieve the strongest classification results.

For this study, we used a 75/25 train/test approach to cross validation. In this approach, we repeated random forest models 1000 times, each using a bootstrapped random sample. Each random sample used 75% of the observations in our dataset to train an algorithm to predict our outcome (using medication for mental health needs). We then used the remaining 25% of the sample (the holdout sample) to test the resultant model on a new sample that had not been tested previously. We repeated this procedure 1000 times and the averaged results are what we report in this article. In each iteration, selection into the train or test condition was stratified on each participant’s SIS Support Needs Index score in order to ensure that the training and testing groups had equivalent representation of participants with varying support needs.

Classification tree models.

A classification tree (CART) predicts an outcome by recursively partitioning the sample into progressively “purer” sub-groups. Starting with the full data set, the algorithm evaluates every candidate predictor and split point to find the one that maximizes separation between those who do versus do not use medication for mental-health conditions (typically by minimizing Gini impurity or entropy). That split creates two branches, and the procedure repeats within each branch until stopping rules are met (e.g., a minimum node size or no further impurity reduction). The terminal “leaf” nodes contain the final class predictions, and the entire tree can be read as a flow-chart of decision rules linking predictor values to medication use.

Random-forest models.

A random forest is an ensemble that combines the predictions of many independent classification trees grown on different bootstrap samples of the data (Breiman, 2001). At each split within every tree, the algorithm considers only a random subset of predictors, injecting additional randomness so that individual trees differ in structure and error patterns. Each tree votes on whether a participant uses medication for mental-health conditions; the forest’s final prediction is the majority vote (or average probability) across all trees. Aggregating many decorrelated trees in this way sharply reduces variance relative to a single tree, yielding a more stable and accurate model while retaining the ability to capture complex, nonlinear relationships in the predictors.

XGBoost models.

Extreme Gradient Boosting builds an ensemble of shallow decision trees, where each new tree is trained to correct the mistakes of the trees already in the model. Specifically, after the first tree makes its predictions, residual errors are computed; the next tree is fit to those residuals, and this process iterates for hundreds of trees. At every step, the algorithm chooses splits that most reduce a differentiable loss function (log-loss in our binary classification context) while applying regularization to prevent overfitting. The final prediction for medication use is a weighted sum of all individual tree outputs, with learning-rate and shrinkage parameters governing how strongly each tree influences the ensemble. This boosting strategy yields high predictive accuracy by sequentially refining the model toward harder-to-predict cases (Chen & Guestrin, 2016).

Measuring variable importance

Although machine learning algorithms typically do not produce interpretive tools like p-values and regression coefficients, they do allow for the estimation of variable importance, which then facilitates the ranking and comparison of the relative importance of the variables in reaching an accurate prediction. In our random forest models, for example, a particular variable’s importance in predicting the outcome (use of medication for mental health conditions) was represented by node purity. Node purity refers to the extent to which splitting a tree on a particular variable decreases the residual sum of squares. For each variable, node purity was calculated based on all splits on that variable across all 1000 trees that we used to construct our final model. Higher values suggest greater variable importance in the prediction of medication use for mental health concerns (Grömping, 2009). Figure 1 shows bar plots representing variable importance with associated node purity for the most important of the 850 variables that were tested in our analysis.

Figure 1.

Variable importance plot for random forest model predicting medication use for mental health.

Results

Outcome variable

The overall aim of this study was to train and test an algorithm to predict our outcome variable, which was based on this question from the background section of the NCI-IDD, as reported by case managers or other support people with access to a person with IDD’s file data: “Takes medication for mood, anxiety, and/or psychotic disorders?” with response options of yes, no, or don’t know. After employing the methods above, our results were as follows.

Descriptive statistics and overall trends by medication status

Table 1 describes the sample, stratified by whether or not respondents were prescribed medication for mental health. Out of the total sample of 2,907 participants, 1,687 (58.0%) were prescribed medication for mental health. Table 1 also includes estimates of the standardized mean difference (SMD) between those who were and were not prescribed medication for mental health, with larger SMDs indicating larger differences between groups. For example, compared to those who were not prescribed medication for mental health, individuals who were prescribed medication for mental health tended to have lower levels of ID.

Comparison of machine learning models

After identifying descriptive trends in the data, we next built a series of predictive models for mental health medication prescription using 847 features (variables) available in the NCI along with related SIS scores. We compared the performance of three models on our training data (n = 2,180). Table 2 includes accuracy, ROC AUC, sensitivity, and specificity for all three models. Our preferred model, the random forest model, had the highest ROC AUC value (0.909), our primary determinant for best model choice. nearly the highest overall accuracy (84.6%), and also had the highest sensitivity (88.7%). The classification tree and boosted regression models had slightly lower sensitivity and ROC AUC values and similar levels of accuracy.

Table 2.

Comparison of model performance for classification tree, random forest, and boosted regression models.

Model	Engine	ROC AUC	Accuracy	Sensitivity	Specificity
Training Data (N = 2,180)
Random Forest	ranger	0.912	84.6%	79.0%	88.7%
Classification Tree	rpart	0.885	85.5%	80.7%	89.0%
Boosted Regression	xgboost	0.911	84.9%	82.8%	87.0%
Testing Data (N = 727)
Random Forest	ranger	0.911	85.7%	78.4%	91.0%

Notes. Both training and testing samples included an identical set of 847 variables from the NCI IPS and SIS. Accuracy = the percentage of all cases that were accurately classified. ROC AUC = Receiver Operating Characteristic Area Under the Curve. Sensitivity = the percentage of all true positives (those who were prescribed medication for mental health) who were correctly classified by the model. Specificity = the percentage of all true negatives (those who were not prescribed meds for mental health) who were correctly classified by the model.

After choosing the random forest model as the preferred option, we then tested the model’s performance on a holdout sample of 25% of the original sample (N = 727). These observations were not included in previous model testing and were prepared according to the same procedures described above in the data preprocessing and feature engineering section. The model performed similarly, with an overall classification accuracy of 85.7%, an ROC AUC value of .911, a sensitivity rate of 78.4%, and a specificity rate of 91.0%.

Variable importance analysis

To interpret the practical significance of the random forest model, we also calculated variable importance measures to determine which variables made the most contribution to the accurate prediction of being prescribed mental health medication. Figure 1 plots the 10 most important variables from our analysis of the testing sample in decreasing order of importance. As described above, variable importance is defined as the relative decrease in accuracy if a given variable is excluded from the model.

Thus, more important variables are associated with larger declines in accuracy, and vice versa. The strongest predictor in our dataset was whether or not a person has a specific diagnosis of a mood disorder, followed by whether a person has a specific diagnosis for anxiety, followed by whether a person takes medication for behavioral challenges, followed by whether a person has a specific diagnosis of a psychotic disorder, followed by whether a person needed support for disruptive behavior, followed by whether a person needs support for behavior that is destructive and harmful to others, followed by whether a person has a specific diagnosis for behavioral challenges, and finally, whether a person has any other disabilities other than ID. The residential setting of the participant (e.g. group home, living independently, living with family) and support needs index score were the ninth and tenth most important variables, respectively. The relative importance of each variable is proportional to the height of the bar, so whether a person has a specific diagnosis for mood disorder, with an importance of approximately 150, is roughly 66% more important to the prediction model as whether a person has a diagnosis for anxiety, which has an importance score of approximately 90. The lower section of Table 1 includes descriptive comparisons of all ten variables, along with their standardized mean differences to aid in interpreting variable importance.

Discussion

Study aim 1: testing machine learning model performance

The first aim of this study was to establish the feasibility of our machine learning approach to accurately predict mental health medication use among adults with IDD. Based on our results, the random forest approach was the most feasible and performed quite well. In our training samples (randomly selected 75% of the total sample in each of 1000 iterations) the algorithm we constructed performed with 90% accuracy, based on ROC AUC. In the independent test sample (a holdout of 25% of respondents in each of 1000 iterations) accuracy in our testing came to 91%. Thus, findings suggest that the use of machine learning to understand factors from the NCI-IDD, SIS, and Medicaid LTSS expenditures that help predict pharmacological interventions for mental health concerns among adults with IDD is feasible. Based on this conclusion, we would be comfortable in suggesting further uses of machine learning to explore personal outcomes for people with IDD in future studies.

Study aim 2: factors contributing to use of medications for MH

As presented in Figure 1, we extracted the most powerful contributors to our algorithm’s performance in order to illuminate the factors that contribute the most to accurate prediction of the use of medications to treat mental health challenges among adults with IDD. Of the top predictors, three of the most powerful were the presence of a diagnosed anxiety disorder, presence of a mood disorder diagnosis, and presence of a diagnosed psychotic disorder. Each of these results stands to reason and suggests that medications are most often being used for their intended purposes. Since prior literature has found that pharmacotherapies for mental health are used both on- and off-label for adults with IDD (McMahon et al., 2020), the finding that anxiety, mood, and psychotic disorders were primary drivers of our predictive algorithm was a positive finding that suggested medication use that generally appeared consistent with prescribing guidelines. There is a role for social workers to intervene in each of these areas as well to help balance medication use with socially-driven interventions, if social work education can improve its incorporation of IDD in the curriculum as suggested by prior studies (e.g. Kim & Sellmaier, 2020; Ogden et al., 2017).

Though most adults with IDD who used medications for mental health management appear to have some form of documented mental health diagnosis, we were surprised to find that some did not, as indicated by the variable no specific mental health diagnosis being among the strongest drivers of our algorithm. Though reasons for this cannot be explained in our application of machine learning, it is plausible that there is some degree of conflation of mental health and challenging behavior among our sample. The fact that taking medication for a behavioral support needs, presence of disruptive behavior, presence of destructive behavior, and having a documented diagnosis related to behavioral support needs are all among the strongest contributors to the accuracy of our algorithm would seem to reinforce this possible explanation for the presence of no specific mental health diagnosis as a strong predictor of medication use. Though the literature has become more clear about the fact that mental health and behavioral support needs are distinct (though related) entities (Broda et al., 2023; Gómez et al., 2021), over the years some practitioners have considered behavioral challenges to be manifestations of mental health or even of IDD itself (Mazza et al., 2020). It is feasible that these perceptions may continue to drive some practitioners to prescribe medications for behavioral or mental health challenges among people with IDD without providing a formal mental health diagnosis, pointing to another area for potential social worker intervention.

Also among the strongest drivers of our algorithm was the Support Needs Index, the composite score of overall need from the SIS. Though the reasoning for this finding cannot be determined in the context of our study, it is plausible that the presence of a mental health condition that necessitates pharmacological intervention may contribute to an overall increase in the amount of support a person with IDD needs to live, work, and play in their community. Further examination of the relationship between support needs, the presence of mental health conditions, and medication use would be a fruitful topic for additional research

It should be noted that most of our strongest predictors of medication use for mental health conditions were either the presence of a particular mental health condition or factors that would logically associate with the presence of a mental health condition, making this study’s findings something of a manipulation check. Future analyses may build on our findings by examining specific treatment approaches or other factors with a more direct line of reasoning to clinical intervention, since variables representing specific clinical approaches are not present in the datasets that were available for our study.

Implications for social work mental health practice

Our descriptive finding that 58% of people with IDD in our multi-year random sample were prescribed at least one medication to treat mental health condition is in line with prior findings (e.g. Iacono & McCarthy, 2020; Kapp et al., 2014) that have consistently shown levels of psychotropic medication use far higher than would be expected in the overall population. This heavy reliance on pharmacotherapies can have detrimental effects and long term side-effects. Though our study cannot make claims about the reasons for such high medication reliance, it is feasible to speculate that absence or shortage of qualified providers to deliver other forms of therapeutic services tailored for people with IDD may play a role in the explanation. As the country’s largest profession within the mental health workforce (U.S. Health Resources and Services Administration, 2023), mental health social workers are in a unique position to serve the needs of adults with IDD who require support for their mental health. Though this study is about the use of medication as one way of treating mental health concerns, it also points to several powerful ways in which social workers may intervene to address mental health concerns and potentially reduce reliance on medications, which are often prescribed for adults with IDD at high doses with risk of serious short- and long-term side effects (Hollins & Attard, 2012). As trained providers of therapeutic services that can help address anxiety, mood, and psychotic disorders, three of the most prominent drivers of pharmacological interventions, based on our findings, mental health social workers may wish to reach out specifically to disability organizations to build partnerships and offer therapeutic services. People with IDD are an underserved population in mental health services, with many mental health service providers not having adequate training to provide service specifically tailored to people with IDD. Therefore, engagement of social workers in more IDD-specific continuing education could help to expand their client bases and prepare them to provide disability competent care. This would be a huge step forward toward reducing not only reliance on pharmacotherapies, but also provide people with IDD skills and strategies they can use to improve their long-term wellness, and hopefully begin to reduce an array of documented health disparities (NIH, 2023).

Beyond the delivery of mental health services, social workers may also work with adults with IDD in case management or other service provision settings. In such contexts, it is important for social workers to be aware of the medications that an adult with IDD is using to treat mental health challenges, to educate about potential side effects, and to link with additional services options as needed. This linkage should include behavior management planning, since our study, as with prior studies (Einfeld et al., 2016) has suggested that mental health medications may be prescribed to treat behavior that can be addressed in non-pharmacological ways if proper supports are in place.

Limitations

Though this study used innovative and advanced methods to examine medication usage for support of mental health needs for adults with IDD, it has limitations that must be noted. The data come from a single state. Though the data we used represented a random sample of adults with IDD who use LTSS in our state, IDD service systems vary from one state to another, so caution should be exercised before assuming that these findings will remain valid in the context of other state service systems. Additional caution should be exercised in extrapolating our results to adults with IDD who do not use LTSS, since their profile may differ somewhat. The dataset contained five years worth of data, which spanned pre-, during, and post-COVID-19 pandemic, which was a volatile time for both IDD service systems and for mental health of people with IDD. We would recommend further analyses in the future, since findings from this volatile time may or may not continue in time of greater stability. Finally, our datasets had constraints that prevented us from asking more nuanced questions about additional therapies received, duration or dosage of medication use, the specific type of pharmacotherapy used, and other factors that may have provided additional useful insight. While such limitations exist in any secondary data analysis, we would encourage future researchers to build such factors into their analysis in order to extend our initial investigation.

Conclusion

Adults with IDD have been shown to exhibit mental health concerns at a rate higher than would be expected in the general public, and are quite likely to use pharmacotherapies for treatment of mental health concerns. Though pharmacotherapies can often be useful as an overall treatment approach, there is also risk of short- and long-term side effects. Our analyses established both that machine learning is a feasible approach for understanding the profile of people with IDD who may be prescribed medication for mental health concerns, and that there are many entry points for mental health social workers to intervene with non-pharmacological interventions. Balanced, person-centered treatment planning is beneficial and necessary, and mental health social workers can be important players in that process.