Predicting Hospital Readmission Rates with Machine Learning

Background to the Study

Hospital readmissions, particularly unplanned readmissions occurring within 30 days of discharge, have emerged over the past two decades as one of the most prominent and actionable indicators of healthcare quality, patient safety, and care coordination effectiveness globally. A readmission signals that a patient's initial hospitalisation was, in some respect, clinically incomplete, or that the transition from hospital to community care was inadequately supported. Preventable readmissions represent a simultaneous failure on three fronts: unnecessary suffering and risk to patients, consumption of scarce hospital bed capacity, and substantial direct and indirect costs to patients, families, and health system administrators.

The global scale of the problem is substantial. Research on Medicare administrative data in the United States found that close to a fifth of Medicare beneficiaries discharged from hospital were readmitted within 30 days, generating tens of billions of dollars in annual readmission-related expenditure. In the United Kingdom, an analysis by the Health Foundation estimated that roughly 30% of emergency hospital admissions were readmissions, at substantial annual cost to the NHS. Across these systems, a meaningful share of readmissions, estimates range widely depending on clinical condition and healthcare system, are potentially preventable through better inpatient care quality, more rigorous discharge planning, and stronger post-discharge support.

Policy responses have been most systematically developed in the United States, where the Affordable Care Act of 2010 established the Hospital Readmissions Reduction Programme, administered by the Centers for Medicare and Medicaid Services, penalising hospitals with excess 30-day readmission rates above risk-adjusted benchmarks for conditions including heart failure, acute myocardial infarction, and pneumonia. Between 2012 and 2019, readmission rates for its target conditions declined by roughly 1.4 to 3.4 percentage points. Full programme details are available through the CMS Hospital Readmissions Reduction Program page.

In sub-Saharan Africa and Nigeria specifically, hospital readmission has received far less systematic research and policy attention than in high-income countries, despite evidence suggesting readmission rates comparable to or exceeding those in Western settings. Research reviewing heart failure management in Mozambican tertiary hospitals reported 30-day readmission rates above 22%, while the multi-country INTER-CHF study spanning heart failure centres across Africa, including Nigeria, reported six-month readmission rates as high as 34.4% for African heart failure patients, substantially exceeding rates in other global regions. These elevated rates are attributed to late presentation with advanced disease, limited access to optimal pharmacological therapy, inadequate post-discharge follow-up infrastructure, patient financial constraints limiting medication adherence, and the absence of systematic risk stratification tools at discharge, concerns that align closely with the priorities the World Health Organization's patient safety programme tracks globally around avoidable harm in health care.

The emergence of predictive modelling as a practical clinical tool for readmission risk stratification represents one of the most significant methodological developments in health services research of the past two decades. Traditional clinical readmission risk scores, including the LACE index and the HOSPITAL score, demonstrated that a small number of routinely available clinical variables could predict readmission with moderate accuracy, with discrimination in the range of 0.68 to 0.74. More recently, machine learning approaches, including Random Forest, Gradient Boosting, and deep learning models applied to electronic health record data, have achieved substantially higher predictive accuracy, with reported discrimination in the range of 0.80 to 0.92 across multiple studies.

Nigeria's tertiary teaching hospitals, including UNTH Enugu, Lagos University Teaching Hospital, Aminu Kano Teaching Hospital, and University College Hospital Ibadan, serve large and complex patient populations who often present late with advanced disease and multiple comorbidities, operating under bed occupancy rates frequently exceeding 100% and fragmented communication between hospital and primary care. The potential value of identifying high readmission-risk patients before discharge is arguably even greater in these settings than in better-resourced Western hospitals. This study addresses the gap between the growing global evidence base on statistical readmission prediction and the near-complete absence of such tools in Nigerian tertiary hospitals by developing, testing, and comparing five competing predictive models using patient data from UNTH Enugu.

Statement of the Problem

Despite the well-documented clinical and economic burden of preventable hospital readmissions in Nigerian tertiary care settings, the development and deployment of statistical predictive models for readmission risk stratification at discharge has received virtually no systematic research attention within Nigeria. Clinicians at UNTH Enugu and peer institutions currently make discharge decisions and post-discharge planning arrangements based exclusively on clinical judgment, experiential heuristics, and available bed pressure, without access to quantitative readmission risk scores derived from local patient data.

This absence of validated predictive tools means high-risk patients, who would benefit from intensified discharge counselling, scheduled follow-up appointments, medication reconciliation, and post-discharge telephone monitoring, are not systematically identified before discharge. Resources for post-discharge support, inherently limited in the UNTH context, are not rationally allocated to the highest-risk patients, producing a cycle of preventable readmissions that burdens patients, families, and the broader healthcare system.

Existing global readmission prediction tools, including the LACE index and HOSPITAL score, were developed and validated on Western patient populations whose clinical characteristics, comorbidity profiles, and disease severity distributions differ substantially from those of Nigerian tertiary hospital patients. Applying these tools without local recalibration risks producing systematically biased predictions that may be more harmful than helpful in the UNTH context.

This study specifically addresses the absence of any validated 30-day readmission predictive model calibrated to the UNTH patient population, the lack of evidence on which clinical and sociodemographic variables are the most powerful readmission predictors in this setting, the absence of comparative evidence on the relative accuracy of traditional statistical versus machine learning predictive approaches using Nigerian hospital data, and the absence of actionable recommendations for integrating readmission risk prediction into UNTH clinical workflows.

Aim and Objectives of the Study

The primary aim of this study is to develop, validate, and compare five statistical and machine learning models for predicting 30-day unplanned hospital readmission using patient discharge data from the University of Nigeria Teaching Hospital, Ituku-Ozalla, Enugu State. The specific objectives are to:

1. Describe the sociodemographic and clinical characteristics of adult inpatients admitted to UNTH Enugu between January 2020 and December 2023.

2. Determine the overall 30-day unplanned readmission rate and its variation by diagnosis category, age group, prior admission history, and discharge disposition.

3. Examine bivariate associations between patient-level factors and 30-day readmission status using chi-square tests and correlation analysis.

4. Develop a Binary Logistic Regression model and identify statistically significant independent predictors of 30-day readmission.

5. Develop Decision Tree, Random Forest, Gradient Boosting Machine, and Lasso penalised logistic regression models for readmission classification.

6. Compare the five models using AUC-ROC, accuracy, precision, recall, and F1-score metrics on a holdout test dataset.

7. Conduct feature importance analysis to identify the most influential predictors across models.

8. Make evidence-based recommendations for integrating readmission prediction into UNTH's clinical workflows and for broader readmission reduction policy in Nigeria.

Research Questions

1. What are the sociodemographic and clinical characteristics of the patient population in the study dataset?

2. What is the overall 30-day unplanned readmission rate, and how does it vary by diagnosis category, age group, number of prior admissions, and discharge disposition?

3. Which patient-level clinical and sociodemographic factors are significantly associated with 30-day unplanned readmission?

4. Which of the five statistical and machine learning models achieves the highest predictive accuracy for 30-day readmission on the holdout test dataset?

5. What are the most influential predictor variables for readmission risk, as identified through feature importance analysis of the best-performing model?

Significance of the Study

This study carries significance on several fronts. Clinically, it provides the first validated readmission prediction model calibrated to the UNTH Enugu patient population, giving clinicians a quantitative tool for discharge risk stratification and specific, actionable levers for readmission prevention, particularly follow-up appointment scheduling and discharge disposition.

Institutionally, a Gradient Boosting Machine-based readmission prediction framework, if integrated into UNTH's clinical information system, offers a mechanism for targeting scarce post-discharge support resources to the highest-risk patients. At the policy level, the study provides empirical evidence supporting a national hospital readmission reduction programme, and the finding that scheduled follow-up appointments significantly reduce readmission risk offers direct policy justification for mandating follow-up scheduling as part of national hospital discharge protocol.

Academically, the study contributes to the still-limited literature on clinical predictive modelling using Nigerian healthcare data, offering a replicable methodological template for similar studies at other Nigerian teaching hospitals. Students working on related quantitative healthcare projects may also find ScholarNest's research coaching service useful for feedback on study design and analysis.

Scope of the Study

This study is geographically limited to UNTH, Ituku-Ozalla, Enugu State. Temporally, it covers inpatient admissions with discharge dates between January 1, 2020, and December 31, 2023, providing 1,200 patient records for analysis. It is restricted to adult patients aged 18 and above admitted to medical and surgical wards; obstetric admissions, same-day surgical cases, psychiatric admissions, and patients who died during the index admission are excluded, as predictors and prevention strategies differ substantially for these groups.

The predictor variables are limited to those routinely captured in the hospital's discharge record system, reflecting what would realistically be available for real-time prediction at discharge. The outcome variable is 30-day unplanned readmission to UNTH only; readmissions to other facilities are not captured and represent a potential source of outcome misclassification.

Operational Definition of Terms

30-Day Unplanned Hospital Readmission: An inpatient hospital admission to UNTH Enugu that occurs within 30 calendar days of discharge from a previous inpatient stay, where the readmission was not planned and recorded as such at the time of the initial discharge.

Charlson Comorbidity Index (CCI): A validated comorbidity scoring system that assigns weighted scores to 17 predefined medical conditions, with higher scores reflecting greater comorbidity burden and higher predicted mortality. In this study, CCI scores are derived from ICD-10 codes recorded in discharge summaries.

Discharge Disposition: The care setting or destination to which a patient is discharged following hospital stay, categorised in this study as home (routine discharge), home with services, discharge against medical advice, and skilled nursing facility or rehabilitation centre.

Gradient Boosting Machine (GBM): An ensemble supervised machine learning algorithm that builds a sequence of decision trees iteratively, with each successive tree trained to correct the prediction errors of the preceding ensemble. The XGBoost implementation is used in this study.

Lasso Logistic Regression: Logistic regression with an L1 (Lasso) penalty term added to the log-likelihood function, shrinking some regression coefficients to exactly zero and thereby performing automatic variable selection, with the penalty parameter tuned via cross-validation.

Random Forest: An ensemble machine learning algorithm that constructs a large number of decision trees during training using bootstrap sampling and random feature selection, and produces predictions by aggregating the predictions of all constituent trees.

Area Under the ROC Curve (AUC-ROC): A model-agnostic performance metric for binary classification models ranging from 0.5, no discrimination, to 1.0, perfect discrimination, equal to the probability that the model ranks a randomly selected positive case higher than a randomly selected negative case.

Feature Importance: A measure quantifying the relative contribution of each input variable to the predictive power of a machine learning model, typically measured as mean decrease in Gini impurity or as SHAP values in Random Forest and GBM.

Index Admission: The initial hospital admission from which readmission status is measured. The 30-day window begins on the date of discharge from the index admission.

Conclusion

Roughly one in five patients readmitted within 30 days is a number that should trouble any hospital administrator, but the more troubling fact is how identifiable that risk turned out to be. Prior admissions, comorbidity burden, and whether a follow-up appointment was scheduled at discharge together explained the majority of predictive power in this study's best-performing model. That is not a call for exotic technology; it is a call for hospitals to systematically capture the data they already generate and use it to flag the patients who need a phone call before they need another admission. Students and researchers exploring related healthcare analytics or statistical modelling questions can find further reference material in the ScholarNest project research library, including comparable studies in applied statistics.

Frequently Asked Questions

Which model predicted hospital readmission most accurately in this study?

The Gradient Boosting Machine achieved the highest performance, with 86.7% accuracy, an AUC-ROC of 0.847, and an F1-score of 0.742, outperforming Random Forest, Lasso logistic regression, standard logistic regression, and Decision Tree.

What was the overall 30-day readmission rate in this study?

The overall 30-day readmission rate was 18.7%, or 224 readmissions out of 1,200 admissions, with a 95% confidence interval of 16.5% to 20.9%.

What are the strongest predictors of hospital readmission?

Number of prior admissions, absence of a scheduled follow-up appointment, discharge against medical advice, and Charlson Comorbidity Index score were the strongest independent predictors identified through logistic regression and confirmed by feature importance analysis.

Which patient group had the highest readmission rate?

Cardiovascular disease patients had the highest readmission rate at 25.8%, followed by endocrine and metabolic conditions including diabetes mellitus at 22.4%.

Why do Western readmission risk tools like the LACE index need local recalibration?

Tools such as the LACE index and HOSPITAL score were developed and validated on Western patient populations whose comorbidity profiles, socioeconomic circumstances, and disease severity differ substantially from Nigerian tertiary hospital patients, so applying them without recalibration risks biased predictions.

What data was used in this study?

The study used de-identified discharge records of 1,200 adult inpatient admissions at the University of Nigeria Teaching Hospital, Enugu, covering January 2020 to December 2023.

What is the Hospital Readmissions Reduction Program?

It is a US Medicare value-based purchasing program administered by the Centers for Medicare and Medicaid Services that financially penalises hospitals with excess 30-day readmission rates for specific conditions.

What practical recommendations does this study make?

It recommends implementing Gradient Boosting Machine-based readmission risk scoring at discharge, mandating follow-up appointment scheduling for high-risk patients, establishing post-discharge telephone follow-up programmes, and developing a national readmission reduction initiative in Nigeria.

What is the Charlson Comorbidity Index?

It is a validated scoring system that assigns weighted scores to 17 predefined medical conditions, with higher scores reflecting greater comorbidity burden and higher predicted mortality risk.

Where can I find more research like this?

Related studies in applied statistics and predictive modelling are available in the Statistics section of the ScholarNest project research library.