Adherence Risk Stratification in Physiatry: A Multivariate Analysis of Factors in Community-Based Care Using Algorithmic Modeling Techniques

Adherence Risk Stratification in Physiatry: A Multivariate Analysis of Factors in Community-Based Care Using Algorithmic Modeling Techniques | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Adherence Risk Stratification in Physiatry: A Multivariate Analysis of Factors in Community-Based Care Using Algorithmic Modeling Techniques Hai Trinh , Robertus Kounang doi: https://doi.org/10.1101/2025.04.26.25326474 Hai Trinh 1 Department of Physical Medicine and Rehabilitation, Arrowhead Regional Medical Center , Colton, CA, USA BA Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: trinhh{at}armc.sbcounty.gov Robertus Kounang 2 Head of Physical Medicine and Rehabilitation MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract Missed appointments represent a double-edged sword in community health settings. Policies designed to retain patients and ensure continuity of care for vulnerable populations often mean that discharging patients is rare, even in cases of frequent no-shows. However, this retention strains healthcare resources, disrupts workflows, and exacerbates inequities in access to care. In physiatry (PM&R), where rehabilitation outcomes depend on consistent patient engagement, missed visits further hinder progress, delaying recovery and diminishing quality of life. Addressing appointment adherence in these settings is paramount for equitable and efficient care delivery. This study evaluated appointment adherence using EPIC-derived general risk scores (demographics, clinical history, and other individual-level factors) and preventative gap scores (compliance with recommended preventative care guidelines). To add granularity, demographic variables such as age, sex, and race/ethnicity—Social Drivers of Health (SDOH) factors embedded within these risk scores—were further analyzed as an additional layer to identify structural and systemic barriers influencing patient engagement. A Residual Deep Neural Network (RDNN) was developed, achieving an AUC-ROC of 0.997, recall of 0.988, F1-score of 0.987, and accuracy of 0.980. A Deep Neural Network with Attention (DNNA) was introduced for interpretability, offering opportunities to refine and extend RDNN’s predictive performance. It demonstrated a 5.7x improvement over a clamped baseline for no-show risk prediction. These findings emphasize the strengths of combining RDNN’s robust predictive capabilities with DNNA’s ability to model nuanced relationships. Together, they provide a pathway to optimize appointment adherence and enhance equitable care delivery in community health and PM&R settings. Introduction Medical no-shows pose operational and clinical challenges, particularly in PM&R, where consistent engagement is critical for rehab outcomes 1 – 3 . Across community health settings, such as those served by Arrowhead Regional Medical Center (ARMC), a public county hospital, policies designed to ensure continuity of care for vulnerable populations often result in retaining patients who frequently miss appointments, further straining workflows and exacerbating inequities in access to care. A small minority of patients often account for a disproportionate share of resource utilization *cite*. These challenges are particularly pronounced in resource-constrained settings, where addressing appointment adherence is critical for both equitable care delivery and operational efficiency. An inherent complexity of the problems is that many elements contribute to this issue, making it difficult for any current model to account for all variables and achieve perfect accuracy 4 – 7 . This study attempts to do that by evaluating the relationship between no-show behavior and predictive factors, such as EPIC-derived general risk scores (encompassing demographics, clinical history, and other individual-level factors) and preventative gap scores (measuring compliance with recommended preventative care guidelines). To add granularity, demographic variables such as age, sex, and race/ethnicity—factors embedded within these risk scores—were analyzed. SDOH indicators were incorporated as an additional layer to identify structural and systemic barriers influencing patient engagement that disproportionately impact PM&R patients in the Inland Empire (IE) 8 – 12 . Two machine learning models—RDNN and DNNA—were utilized to predict no-show risks. The RDNN demonstrated exceptional reliability, achieving an AUC-ROC of 0.997, accuracy: 0.980, sensitivity: 0.988, precision: 0.987, and F1: 0.987. Complementing RDNN, DNNA offered potential enhancements by capturing nuanced feature interactions and achieving a 5.7x improvement over a clamped baseline for no-show prediction. Together, these models provide actionable insights to optimize appointment adherence in a resource-constrained environment, enabling targeted interventions that can improve patient engagement and care continuity. By highlighting the critical role of predictive factors such as general risk and preventative gap scores, this work demonstrates how machine learning can uncover meaningful patterns and support targeted interventions. These findings underscore the potential for integrating advanced machine learning models into community health and PM&R settings to enhance equitable care delivery and patient outcomes. Methods Study Design & Setting This retrospective cohort analysis utilized de-identified EHR data from EPIC, spanning five years (Jan 1, 2020 – Jan 1, 2025) , to evaluate the impact of SDOH on no-show rates in PM&R. Patient Population A total of 968 patients met the criteria. Inclusion Criteria Adults ≥30 yrs : This threshold was selected to capture a population where career stability, functional limitations, and chronic disease burden are more pronounced, creating a more stable baseline for evaluating SDOH influences on no-show rates while reducing variability from transient life-stage factors. Exclusion Criteria Patients <18 yrs: Excluded due to distinct clinical and logistical factors, including developmental differences, parental dependency, and school-related scheduling conflicts, which may introduce confounding influences unrelated to individual-level SDOH. Patients 18-29 yrs : Excluded as this age group often experiences similar education level, employment history, and generally insurance coverage under parental plans, creating greater homogeneity within the dataset and potentially obscuring patterns in predictive scores. Patients in acute rehab (<30 days) : Excluded because adherence in short-term institutionalization is dictated by facility protocols rather than individual behavior or predictor variables. Dataset overview No-show frequency : Captured as a binary outcome (0 = no missed appointments, 1 = at least one missed appointment) Key demographics : Age, legal sex, OMB ethnicity and race. General Risk Score : An EPIC-derived metric based on patient medical history (chronic conditions, prior hospitalizations, and comorbidities), vital signs (e.g., blood pressure and heart rate), and laboratory results (key lab values relevant to the patient’s health status). Preventive Care Gap Score : Reflects compliance with recommended preventive care, including adherence to routine screenings, vaccination schedules, and chronic disease management guidelines. 20 SDOH Domains : Alcohol Use Depression Digital Connectivity Financial Resource Strain Food Insecurity Health Literacy Homeless Housing Stability Intimate Partner Violence Medical Cost Burden Personal Safety Physical Activity Postpartum Depression Safety and Environment Social Connections Stress Tobacco Use Transportation Needs Utilities None of the Above Model design RDNN : Produced metrics with higher predictive accuracy compared to other models such as XGBoost and Random Forest. DNNA : Enhanced RDNN’s performance by demonstrating a 5.7x improvement over a clamped baseline, enabling deeper analysis of relationships between risk factors. Evaluation Metrics AUC-ROC : Assessed the model’s ability to distinguish between patients likely to no-show and those likely to attend, providing an overall measure of classification performance across thresholds. Accuracy : Measured the percentage of total predictions that were correct, offering a straightforward metric of performance across both classes. Sensitivity : Evaluated the model’s ability to correctly identify patients who were no-shows, focusing on minimizing missed high-risk cases. Precision : Measured the proportion of correctly predicted no-shows out of all predicted no-shows, highlighting the model’s ability to minimize false positives. F1 : Balanced precision and sensitivity, serving as a comprehensive measure for models dealing with imbalanced classes. Analysis Framework Models trained and validated using 10-fold cross-validation to ensure generalizability. Feature importance analysis performed to assess contributions of general risk scores, preventive care gap scores, and SDOH domains to no-show predictions. Results stratified by demographic and risk group categories to identify population-specific patterns. Download figure Open in new tab Figure 1: Model Workflow Blueprint Missing values, including care gap and preventive gap scores, were filled using the median value, with binary indicators added to flag missing data. Numeric and ordinal features were scaled between -1 and 1 using the Smooth Ridit transformation. The preprocessed data was then fed into the Keras Residual Deep Neural Network (RDNN) for predictions. Download figure Open in new tab Figure 2: Lift Chart Compare the model’s predicted likelihood of no-shows to the actual outcomes, grouped into bins (ranges of predicted probabilities, such as 0–10%, 10–20%, etc.), to evaluate accuracy and reliability at different levels of risk. View this table: View inline View popup Download powerpoint Table 1. demonstrates the relative importance of features in predicting no-shows. While “General Risk Score” and “Preventive Care Gap Score” incorporate demographic and SDOH factors, the machine learning model identified “Age” as the most influential standalone feature. This result likely reflects the direct and indirect role age plays in healthcare engagement. Age is often a proxy for health status, chronic conditions, and scheduling predictability, all of which influence appointment adherence. Machine learning models, such as the RDNN used here, are designed to optimize predictive accuracy by weighing the contribution of each feature to the target outcome. “Age” emerged as the most impactful feature because it interacts strongly with other predictors and directly correlates with patterns of healthcare utilization. For example, older individuals may face more barriers, such as mobility challenges or comorbidities, increasing the likelihood of no-shows. Younger individuals, on the other hand, might be influenced by transitional life stages or competing priorities. By disentangling individual features like “Age” from composite scores, the model ensures that the impact of key demographic variables is not diluted within broader risk metrics. This dual-layer approach highlights “Age” as both a direct and indirect driver of no-show behavior. Download figure Open in new tab Figure 3: Model Performance Metrics and ROC Curve This figure demonstrates the RDNN model’s exceptional performance in predicting no-shows, achieving an AUC of 0.997 , indicating near-perfect discrimination between no-shows and attended appointments. The confusion matrix reveals a sensitivity of 0.988 and precision of 0.987 , showcasing the model’s ability to accurately identify no-shows while maintaining low false positive rates. With an overall accuracy of 0.980 , the model delivers highly reliable predictions. Results The ROC curve underscores the model’s superior trade-off between sensitivity and false positive rate compared to random classification. The density plot highlights the clear separation in predicted probabilities, with the decision threshold (0.5) effectively distinguishing between no-shows and attendees. These results validate the RDNN’s robustness and practical value in optimizing appointment adherence. Download figure Open in new tab Figure 4: Age Feature Effects The trend shows fluctuations in no-show rates across age groups, with no clear pattern of adherence tied to age. Partial dependence isolates the impact of age on no-show rates by holding other variables constant, providing a clear view of age’s independent effect. For example, if the model evaluates patients aged 50, it assumes all other factors (e.g., risk score, preventive care gaps) remain unchanged. This allows the model to estimate how being 50 alone influences no-show likelihood, without interference from other variables. Actual values represent real-world observations, and the variability across all metrics suggests age alone does not drive adherence trends. The model ranked age as having the highest feature importance (1.0) because it predicts no-shows well across the dataset, largely through its interactions with other features and its role in splitting data for predictions. However, the graph reflects only the isolated effect of age, which fluctuates significantly, indicating that age’s influence is stronger when combined with other factors. Download figure Open in new tab Figure 5: General Risk Score Feature Effects The model identifies general risk score as a strong predictor because it captures the cumulative burden of barriers, such as transportation challenges or chronic conditions. The peak at 6 may indicate a tipping point where these issues significantly disrupt attendance. The dip at 8 might reflect qualifying for programs like Social Security or Medicaid as evident by corresponding improved adherence on the age chart at 65 years old. However, the rise at 10 highlights that the partial dependence curve aligns with actual and predicted values up to 10+, as the model’s predictions reflect the isolated contribution of general risk score. Beyond 10, the divergence likely occurs due to sparse high-risk data, forcing the model to rely on interactions with other variables (e.g., age, interventions) and increasing uncertainty. Download figure Open in new tab Figure 6: Preventive Care Gap Score Feature Effects This plot demonstrates the inverse relationship between preventive care gap scores and no-show rates showing an overall trend of decreasing no-show rates as the gap widens from 0 to 6, potentially reflecting patient motivation to address unmet care needs. The drop at 7 matches the age chart (adherence improves at 65+) and the general risk score chart (score 8 aligns with Medicaid or Social Security access), reflecting most likely addressing the cumulative gaps. At 8 and 9, partial dependence returns to baseline trends potentially due to those who did not receive safety nets or accumulation of burdens became overwhelming. However, predicted and actual values drop sharply, reflecting sparse data and the isolated impact of preventive care gap score without interactions, such as age or general risk score. It may also be that real-world effects are all three figures are more severely skewed (left *insert correct stat term*) for negative outcomes. Download figure Open in new tab Download figure Open in new tab Figure 7: Coefficients of feature effects The magnitude of coefficients reflects the strength of each feature’s contribution, but interactions between features play a critical role in RDNNs. Unlike linear models, RDNN coefficients reflect complex interactions that may not represent isolated feature impacts but rather their contribution to patterns learned by the model. Negative coefficients, like Age-65 and SDOH-High Risk Financial Resource Strain, reflect reduced no-shows, likely due to the patient population served at PM&R. Many patients on Medi-Cal, Medicaid, Workers’ Compensation, Social Security, or disability payments may have stronger incentives to adhere to follow-ups for eligibility extensions. These coefficients highlight not just feature contributions but also the structural incentives shaping adherence within a specific care system. Positive coefficients reflect features that increase no-show likelihood, such as Preventive Care Gap Score = 0, which may instead reflect a population with sufficient access to care who perceive fewer immediate benefits from follow-ups. This group might lack the security-driven incentives present in high-risk populations, aligning with a drop in adherence at age 63, where patients may gauge the likelihood of future benefit approvals, leading to disengagement from services. RDNNs model complex interactions and prevent vanishing gradient problems, where earlier layers struggle to learn because updates get smaller with each step backward through the network, reducing their ability to improve learning. Residual connections solve this by skipping layers, ensuring smooth learning and stable training, while preserving intricate patterns. However, the model treats all features relatively equally, rather than emphasizing those most relevant to specific interactions or contexts. which may limit interpretability as we had to hedge why and how a feature was weighted which for implementation is at times more important than the accuracy of predictability. To address this, a DNNA was deployed to dynamically weight critical features and their interactions to provide a more granular understanding of patient dynamics. This shift allows the model to better prioritize contextual dependencies and feature hierarchies, offering clearer insights for reducing no-show rates and tailoring interventions to specific patient groups. Download figure Open in new tab Figure 8: The “Likely Segment” represents a group of patients identified through clustering, which groups individuals with similar characteristics. A “no-show? 1” indicates these patients had a positive no-show outcome. The segment size reflects the percentage of patients who share overlapping features, without requiring all traits simultaneously. This overlap highlights patterns in shared barriers or risks rather than strict one-to-one feature matching. Segment Rate shows how often no-shows happen in this group, Base Rate shows how often they happen in the entire dataset, and Rate vs. Baseline compares this group to the overall average. Under the RDNN model, patients with SDOH = None of the Above are more likely to adhere to treatment, possibly reflecting fewer reported barriers or underreporting. Conversely, a General Risk Score of 1 , while indicating lower risks, correlates with less appointment adherence, likely due to perceived non-urgency in healthcare needs. However, preventive Care Gap Scores , unlike self-reported SDOH or general risk classifications, offer more objective insight into patient behavior. In this context, where other features like age groups or SDOH domains provide granularity, the preventive care score reveals patterns of disengagement tied to neglect rather than intentional action. This aligns with broader patterns of no-shows, suggesting reluctance or avoidance in engaging with preventive care. Notably, this demographic—older white males with underreported barriers—aligns with broader health and behavioral risk literature, highlighting vulnerabilities such as social isolation, healthcare avoidance, financial strain, and stigma, which encompasses a range of negative health outcomes such as suicide risk ***. Transitioning to a DNNA model can uncover these nuanced interactions and identify tailored interventions to address these systemic and behavioral challenges. Discussion Interpreting the Drivers of No-Show Behavior in PM&R The RDNN achieved strong predictive accuracy (AUC-ROC of 0.997), illustrating how the distinctive characteristics of a community health setting in PM&R can be leveraged to model no-show behavior. Moreover, our DNNA model outperformed a clamped baseline by 5.7x, leveraging its dynamic weighting of feature interactions to more precisely target high-risk segments allowed a more nuanced understanding of patient engagement in PM&R. While Age and SDOH are embedded within the General Risk Score and Preventive Care Gap Score, our analysis revealed that separating these variables is critical for uncovering subpopulations—such as older white males with under self-reports of issues—who might otherwise remain undetected as high risk. Through multivariate analysis, patterns and stories emerge that may remain untold or validated in both specific and general contexts when variables are only viewed in isolation or through composite measures. The overall finding generally aligns with established research suggesting that older adults may experience specific barriers—mobility issues, comorbidities, or transportation limitations—that increase no-show risk ***. However, despite the IE being one of the worst regions in America for transportation infrastructure and public transportation access, Transportation Needs—even for those who self-report as high risk—did not demonstrate significant predictive power. This suggests that current interventions at ARMC: PM&R may be effectively mitigating barriers for patients with functional limitations. Moreover, certain older adults in underserved settings may also be more motivated to attend appointments due to the need for continued eligibility in programs such as Medicaid, Workers’ Compensation, or Social Security. This was evident not only when in terms of a patient’s age where appointments are more likely to be adhered to but that those at high-risk for financial strain were less likely to no-show. These structural incentives can drive more consistent follow-up, especially in public county hospital settings, where vulnerable patients are retained in an effort to promote equitable care and rehabilitation outcomes. Health Literacy—has been repeatedly shown to influence diverse outcomes such as teen pregnancy rates, workforce participation, access to education, and overall quality of life 13 – 15 . Our model reflects this broader evidence of higher rates of no-shows in patients with ‘High Risk’ of Health Literacy. These findings emphasize the need for improved assessment and integration of Health Literacy to enhance predictive accuracy and address systemic gaps in care as a high-yield focus area. Particularly in a public county hospital setting, often intersect with patients’ economic, educational, and societal contexts in ways that composite risk scores alone cannot fully elucidate. Moreover, Caregiver Support frequently goes unreported or functions informally for patients. While “None of the Above” for SDOH may be a protective factor overall, it could equally represent a data-collection error such as older white males or a systemic gap in PM&R. Transportation Needs had one of the largest sample sizes before exclusion (>1,000), while Caregiver Support was underrepresented (<200 patients), limiting its direct influence despite its critical role in PM&R patient adherence. Many patients rely on informal caregivers for scheduling, transportation, and medical decision-making, yet this dependency is underreported in structured SDOH surveys despite frequent self-reporting in clinical interactions. This discrepancy suggests a fundamental gap between what patients disclose in practice and what is systematically captured in EHR data. This underreporting can artificially reduce the predictive importance of SDOH in quantitative models, even though these factors critically shape patient behavior in day-to-day clinical practice. Future Directions Future research should incorporate longitudinal surveys and structured interviews with high-risk patients 16 and those flagged as potential no-show risks based on inconsistent quantitative data trends. Using both structured and open-ended questions, these qualitative methods can elicit patient-reported logistical barriers that structured data often fails to capture—such as a newly developed mobility limitation affecting transportation access or a patient with a family history of diabetes whose adherence patterns are shifting due to increasing caregiving responsibilities. Applying natural language processing (NLP) techniques to analyze patient responses could uncover patterns in reasoning, cognitive load, and self-reported problem-solving abilities that influence adherence 16 . Beyond sentiment and keyword analysis, NLP could assess response complexity, hesitation markers, or conceptual understanding when patients describe scheduling challenges or define key healthcare concepts (e.g., how they perceive “follow-up care” or “appointment confirmation”). By integrating these linguistic indicators with patient-reported barriers, NLP could help refine risk stratification models, distinguishing between patients who lack access and those who may require additional support in navigating healthcare systems. Combining these qualitative insights from a large language model with rule-based predictive algorithms could transform how missing data is interpreted—differentiating passive non-response from active disengagement—allowing for more precise and patient-specific intervention strategies. Limitations Retrospective Design & Causality Constraints As a retrospective study, even with RDNN and DNNA, cannot establish causality between feature factors and no-show behavior. Findings should be interpreted as associations rather than definitive causal effects. Distribution Skew & Influence of Repeat Offenders This pattern suggests a highly skewed distribution, where a small subset of patients (∼45%) accounted for 95% of all no-shows, including extreme cases such as a patient with 41 missed visits. While this study focuses on broader trends, future research should investigate individual-level drivers of extreme no-show frequencies to better inform targeted interventions. A more detailed patient narrative is provided in Appendix A (if journal allows). Generalizability Across Healthcare Settings This study was conducted in a safety-net PM&R setting, where most patients were insured through Medi-Cal, Medicaid, or Workers’ Compensation, reducing direct financial barriers to care. Unlike in private practices, where out-of-pocket costs and cancellation fees incentivize adherence, community-based settings may lack immediate financial consequences for missed appointments.. Small Sample Size and Uniformity While the model is more reliable, ethnicity dropped from a top three predictor in earlier models to a non-significant factor. Although this version improves predictive accuracy, it still lacks the granularity needed to fully capture how sociocultural, geographic, and generational differences influence healthcare access and engagement. The small sample size further limited statistical reliability, particularly for AANHPI patients, where fewer than 30 individuals met the inclusion criteria. Racial and ethnic categories were aggregated to reduce statistical noise, but this approach masked key distinctions. For example, Korean immigrants from urban centers and Indian patients transitioning from rural healthcare systems experience healthcare differently. Similarly, Tejanos relying on VA healthcare and Cuban refugees from state-run systems navigate centralized care in distinct ways, yet their shared Catholic faith may shape similarperspectives on medicine. These lived experiences influence trust, engagement, and utilization, underscoring the limitations of broad racial categorization in predictive modeling. Conclusion This study underscores the precision and power of machine learning models in predicting no-show behavior within the nuanced context of PM&R in the Inland Empire. The RDNN achieved near-perfect performance, with an AUC-ROC of 0.997, an accuracy of 0.980, and an F1-score of 0.987, while the DNNA added depth by dynamically weighting feature interactions, outperforming a clamped baseline by 5.7x. Together, these models provide a confirmatory framework for understanding and addressing appointment adherence challenges in a resource-constrained, community health setting. The General Risk Score and Preventive Care Gap Score, as composite measures, combine multiple embedded variables such as demographics and SDOH. While these composites offer a high-level predictive capacity, this study suggests that by splicing these metrics and analyzing their individual components, we can enhance interpretability 4,7,17 . The disaggregation of Age and SDOH revealed insights into hidden risks and systemic barriers—patterns that would otherwise remain obscured in aggregate analyses. This methodology demonstrates how combining composite scores with a more granular approach can yield a richer understanding of no-show behavior and drive targeted, effective intervention The unique makeup of community health in a PM&R setting within the IE served as an advantage for these models 8,12,19,20 . The interplay of demographic, socioeconomic, and structural factors within this resource-constrained environment created a distinctive ecosystem that amplified the accuracy and applicability of our predictive framework 11 . Embedded incentives like Medicaid and Social Security eligibility further shaped patient behavior, offering a rich context to identify and address no-show risks effectively. In a hyper car-dependent region with a relatively high cost of living compared to the rest of the country, a functionally limited patient population faces challenges in gaining employment and competing in the open labor market. This context helps filter whether no-show risk is primarily driven by car ownership, socioeconomic constraints, access limitations, or age-related mobility issues, connecting these systemic barriers to broader patterns of disengagement 8,9,12,18 . While advanced composite metrics, provided by comprehensive platforms like EPIC, offer valuable insights, they alone are insufficient to fully capture the complexity of patient behavior. However, these platforms provide a strong foundation that, when refined through advanced machine learning models, can significantly enhance interpretability and predictive power by uncovering hidden patterns and high-risk populations that might otherwise remain obscured. This combination of advanced algorithms and robust data systems enables a more nuanced and actionable understanding of patient disengagement. This approach not only demonstrates the robustness of these tools but also underscores the necessity of combining composite metrics with granular, patient-level analysis. By segmenting and integrating variables rather than treating them as standalone factors, the models deliver a more complete and actionable narrative about the drivers of adherence behavior. Future research should build on this methodology by incorporating longitudinal data and patient-reported barriers to further refine the predictive power of these models. These findings pave the way for tailored interventions that improve care delivery and equity, offering a replicable model for leveraging advanced analytics to enhance operational efficiency and patient outcomes in resource-limited healthcare systems. Data Availability Researchers may request access to the analytic code or variable definitions used in this study, pending institutional review and data use agreement. Reference List ↵ Françesc Medina-Mirapeix , Medina-Mirapeix F , Silvana Loana de Oliveira-Sousa , et al. Continuity of care in hospital rehabilitation services: a qualitative insight from inpatients’ experience . Rev Bras Fisioter . 2017 ; 21 ( 2 ): 85 – 91 . doi: 10.1016/j.bjpt.2017.03.002 OpenUrl CrossRef Elisabeth L Rosenthal . Continuity of Care in Rehabilitation . Am J Phys Med Rehabil . 2023 ; 102 ( 2S ): S10 – S12 . doi: 10.1097/phm.0000000000002144 OpenUrl CrossRef PubMed ↵ Merethe Hustoft , Hustoft M , Eva Biringer , et al. The effect of team collaboration and continuity of care on health and disability among rehabilitation patients: a longitudinal survey-based study from western Norway . Qual Life Res . 2019 ; 28 ( 10 ): 2773 – 2785 . doi: 10.1007/s11136-019-02216-7 OpenUrl CrossRef PubMed ↵ Wyatt P. Bensken , Ryan McGrath , Rachel Gold , Erika Cottrell . Area-Level Social Determinants of Health and Individual-Level Social Risks: Assessing Predictive Ability and Biases in Social Risk Screening . J Clin Transl Sci . Published online November 10, 2023:1-24. doi: 10.1017/cts.2023.680 OpenUrl CrossRef Steven Rothenberg , Steven Rothenberg , Bill Bame , Bill Bame , Ed Herskovitz , Ed Herskovitz . Prospective Evaluation of a Machine-Learning Prediction Model for Missed Radiology Appointments . J Digit Imaging . Published online June 29, 2022. doi: 10.1007/s10278-022-00670-3 OpenUrl CrossRef Shahabeddin Sotudian , Shahabeddin Sotudian , Aaron Afran , et al. Social determinants of health and the prediction of missed breast imaging appointments . BMC Health Serv Res . 2022 ; 22 ( 1 ): 1454 – 1454 . doi: 10.1186/s12913-022-08784-8 OpenUrl CrossRef PubMed ↵ Luiz Henrique Américo Salazar , Luiz Henrique Salazar , Parreira W , et al. No-Show in Medical Appointments with Machine Learning Techniques: A Systematic Literature Review . Information . 2022 ; 13 ( 11 ): 507 – 507 . doi: 10.3390/info13110507 OpenUrl CrossRef ↵ Dorlins A. Villalobos , Carolina Carlos , Yasamin Rasouli , Jordan M. Leffew . Balancing Housing Affordability and Transportation Efficiency in the Inland Empire . Int Conf Transp Dev 2024 . Published online 2024. doi: 10.1061/9780784485521.022 OpenUrl CrossRef Biggs J , Njoku N , Kurtz K , Omar A. Decreasing Missed Appointments at a Community Health Center: A Community Collaborative Project . J Prim Care Community Health . 2022 ; 13 : 21501319221106877 . doi: 10.1177/21501319221106877 OpenUrl CrossRef PubMed Lydia Smeltz , Lydia Smeltz , Kheng Hock Lee , Kheng Hock Lee , Janet Prvu Bettger , Janet Prvu Bettger . Integrating rehabilitation and social care: opportunities for advancing research, training, practice, and policy . Arch Phys Med Rehabil . Published online May 1, 2022. doi: 10.1016/j.apmr.2022.04.015 OpenUrl CrossRef Rosalynn R.Z. Conic , Conic RRZ , Carolyn Geis , Geis C , Heather K. Vincent , Vincent HK . Social Determinants of Health in Physiatry: Challenges and Opportunities for Clinical Decision Making and Improving Treatment Precision . Front Public Health . 2021 ; 9 . doi: 10.3389/fpubh.2021.738253 OpenUrl CrossRef ↵ Kimberly Collins , Raffi Der Wartanian , Preston L. Reed , Holly Chea , Yunfei Hou , Yongping Zhang . Social equity and public transit in the inland empire: Introducing a transit equity analysis model . Transp Res Interdiscip Perspect . 2023 ; 21 : 100870 – 100870 . doi: 10.1016/j.trip.2023.100870 OpenUrl CrossRef ↵ Éva Marjorie Couture , Couture ÉM , Maud-Christine Chouinard , et al. The relationship between health literacy and quality of life among frequent users of health care services: a cross-sectional study . Health Qual Life Outcomes . 2017 ; 15 ( 1 ): 137 – 137 . doi: 10.1186/s12955-017-0716-7 OpenUrl CrossRef PubMed Mengyun Zheng , Zheng M , Hui Jin , et al. The relationship between health literacy and quality of life: a systematic review and meta-analysis . Health Qual Life Outcomes . 2018 ; 16 ( 1 ): 1 – 10 . doi: 10.1186/s12955-018-1031-7 OpenUrl CrossRef PubMed ↵ Zhang Q. The cost of illiteracy: A causal inference study on how illiteracy affects physical and mental health . Health Educ J . 2021 ; 80 ( 1 ): 54 – 66 . doi: 10.1177/0017896920949894 OpenUrl CrossRef David A. Dorr , David A. Dorr , Ana Quiñones , et al. Prediction of Future Health Care Utilization Through Note-extracted Psychosocial Factors . Med Care . 2022 ;Publish Ahead of Print. doi: 10.1097/mlr.0000000000001742 OpenUrl CrossRef Kevin P. Fiori , Fiori K , Caroline Heller , et al. Unmet social needs and no-show visits in primary care in a US Northeastern urban health system, 2018–2019 . Am J Public Health . 2020 ; 110 . doi: 10.2105/ajph.2020.305717 OpenUrl CrossRef Yi Yang , Samaneh Madanian , David Parry . Enhancing Health Equity by Predicting Missed Appointments in Health Care: Machine Learning Study . JMIR Med Inform . Published online 2024. doi: 10.2196/48273 OpenUrl CrossRef Kimberly Collins , Raffi Der Wartanian , Francisca Beer , Yunfei Hou . Moving Towards the Electrification of Medium- and Heavy-Duty Vehicles in the Inland Empire . Published online 2024. doi: 10.31979/mti.2024.2305.1 OpenUrl CrossRef Jai Radhakrishnan , Jai Radhakrishnan , Martin Collazo , et al. Towards Predicting Bus On-Time Performance in the Inland Empire . 2021 IEEE Int Conf Big Data Big Data . Published online December 15, 2021. doi: 10.1109/bigdata52589.2021.9671477 OpenUrl CrossRef View the discussion thread. Back to top Previous Next Posted April 28, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Adherence Risk Stratification in Physiatry: A Multivariate Analysis of Factors in Community-Based Care Using Algorithmic Modeling Techniques Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Adherence Risk Stratification in Physiatry: A Multivariate Analysis of Factors in Community-Based Care Using Algorithmic Modeling Techniques Hai Trinh , Robertus Kounang medRxiv 2025.04.26.25326474; doi: https://doi.org/10.1101/2025.04.26.25326474 Share This Article: Copy Citation Tools Adherence Risk Stratification in Physiatry: A Multivariate Analysis of Factors in Community-Based Care Using Algorithmic Modeling Techniques Hai Trinh , Robertus Kounang medRxiv 2025.04.26.25326474; doi: https://doi.org/10.1101/2025.04.26.25326474 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Rehabilitation Medicine and Physical Therapy Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (299) Cardiovascular Medicine (4422) Dentistry and Oral Medicine (443) Dermatology (382) Emergency Medicine (607) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1507) Epidemiology (15219) Forensic Medicine (30) Gastroenterology (1123) Genetic and Genomic Medicine (6584) Geriatric Medicine (667) Health Economics (997) Health Informatics (4524) Health Policy (1367) Health Systems and Quality Improvement (1611) Hematology (540) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15910) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (145) Nephrology (667) Neurology (6587) Nursing (345) Nutrition (998) Obstetrics and Gynecology (1143) Occupational and Environmental Health (956) Oncology (3331) Ophthalmology (970) Orthopedics (369) Otolaryngology (420) Pain Medicine (435) Palliative Medicine (129) Pathology (663) Pediatrics (1690) Pharmacology and Therapeutics (691) Primary Care Research (710) Psychiatry and Clinical Psychology (5437) Public and Global Health (9218) Radiology and Imaging (2195) Rehabilitation Medicine and Physical Therapy (1369) Respiratory Medicine (1195) Rheumatology (593) Sexual and Reproductive Health (709) Sports Medicine (529) Surgery (709) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ff79574a911e2c5',t:'MTc3OTQwODQ3OQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();