TL;DR
This paper tackles the critical issue of predicting unplanned hospital re-admissions for patients with alcohol use disorders (AUD) within 28 days. By employing a suite of machine learning models, the researchers identified key patient characteristics and post-discharge interactions that influence re-admission risk, with the Clinical Bio-BERT model showing the highest sensitivity for prediction.
What Problem Does This Paper Address?
Patients with alcohol use disorders (AUD) face a significantly higher risk of hospital re-admission, placing a substantial burden on healthcare systems. These unplanned re-admissions often signal suboptimal care and are costly. While previous research has identified some risk factors, there's a need for more robust predictive models to enable targeted interventions and reduce these readmissions. This study aims to fill this gap by developing and evaluating machine learning approaches to predict 28-day all-cause unplanned hospital re-admissions specifically for AUD patients in an Australian regional health district.
Methods
The study employed a retrospective design, utilizing de-identified linked datasets from a local health district. After data preprocessing, the researchers developed five machine learning models to predict 28-day re-admission:
- Logistic Regression (LR): Used as a baseline model.
- Random Forest (RF): An ensemble learning method that builds multiple decision trees.
- Support Vector Machine (SVM): A model that finds an optimal hyperplane to separate data points.
- Long Short-Term Memory (LSTM): A type of recurrent neural network well-suited for sequential data, like time-series clinical records.
- Clinical Bio-BERT: A transformer-based model pre-trained on clinical text, adapted for this task.
The models were trained and tested using various validation techniques, including 10-fold cross-validation for RF and SVM, and a 70/30 train-test split for LSTM and Clinical Bio-BERT. Model performance was evaluated using metrics like Area Under the Curve (AUC), accuracy, specificity, and sensitivity.
Datasets
The study utilized four de-identified datasets from the Illawarra Health Information Platform:
- Emergency Department (ED) data
- Hospital Admitted Patient (AP) data
- Community-based Drug and Alcohol Service (D&A service) data
- Mental Health Service data
These datasets were linked by unique patient identifiers. The final analysis included 869 patients with alcohol use disorders who incurred 2254 hospital admissions between 2015 and 2018. Patients with interactions with mental health services were excluded to improve data homogeneity.
Key Insights and Ideas
- Post-discharge engagement is protective: Interactions with the Emergency Department (ED) or Drug and Alcohol (D&A) services after discharge significantly reduced the risk of re-admission. This highlights the crucial role of continued care and support networks.
- Specific patient profiles are at higher risk: Certain demographic and clinical characteristics are strong predictors of re-admission, allowing for early identification of high-risk individuals.
- Machine learning models offer varying strengths: While LSTM showed high overall performance, Clinical Bio-BERT demonstrated the highest sensitivity, making it a promising tool for identifying patients most likely to be re-admitted.
- Comorbidities matter, but the nuance is complex: While certain comorbidities (respiratory, hepatobiliary) increased risk, the study found a counterintuitive finding regarding the "complications and comorbidities" variable, suggesting the need for further investigation into how severity is measured and interpreted.
Key Results
- Risk Factors Identified: Ten variables were identified as significant predictors of 28-day re-admission. These include:
- Age groups: 45-49, 70-74, and 75-79 years old were significantly more likely to be re-admitted.
- Sex: Males had a 36% higher risk than females.
- Polysubstance use: Patients using substances in addition to alcohol were over 3 times more likely to be re-admitted.
- Specific Diagnoses: Respiratory system disorders and hepatobiliary system and pancreas disorders were associated with a ~1.6 times higher risk.
- Post-discharge Interactions: Interactions with ED or D&A services reduced re-admission risk by 71% and 79%, respectively.
- Model Performance:
- The 10-variable LSTM model achieved the highest accuracy (0.814) and a strong AUC of 0.749.
- The 42-variable Clinical Bio-BERT model achieved the highest sensitivity (0.724) and an AUC of 0.753.
Limitations and Future Work
The authors acknowledge several limitations:
- Data Restrictions: The study was limited to a single health district, potentially limiting generalizability. External validation with data from diverse settings is recommended.
- D&A CL Service Data: Data on the specific impact of hospital-based D&A consultation liaison (CL) services was unavailable, hindering a complete assessment of D&A service effectiveness.
- Mental Health Exclusions: Patients with co-occurring mental health issues were excluded, limiting the study's scope for dual-pathology patients. Future research should investigate risk factors for this group.
- Comorbidity Variable Definition: The lack of clear definitions for the "complications and comorbidities" variable limits interpretability. Further research is needed to understand how cumulative effects impact re-admission.
- Social Determinants of Health: Variables like socioeconomic status, education, and housing were not included due to data unavailability. Incorporating these could enhance predictive accuracy and real-world applicability.
- Hospital Mortality: The study did not account for hospital mortality in its re-admission calculations.
Future work should focus on external validation, incorporating unstructured data into BERT models, investigating dual-pathology patients, and exploring the impact of social determinants of health.
Summary updated