Operational Readiness

Operational Readiness You’ve been hired as a data analyst at a mid-sized hospital exploring AI models to reduce 30-day readmission rates. Reflect on the importance of model selection and evaluation. Which algorithms would you prioritize and why? How would you ensure leadership understands both the benefits and limitations of your recommendation?

Using AI systems to reduce 30-day readmission rates is paramount because it helps to achieve higher-quality care, better patient outcomes, lower healthcare expenses, lower mortality risk, financial incentives, and better public health… in a market such as the U.S. where the healthcare costs grow higher every year. Before choosing a model(s), understanding the severity of the causes is essential. The following are the top 5 factors causing patient readmissions [1]:

1. Non-compliance/lack of follow up: inability or unwillingness to take medication correctly due to cost, confusion, or management difficulties. Missing post-discharge appointments and ending up in the E.R. to be readmitted for conditions that could have been managed in an outpatient setting.
2. Medication issues: inaccurate medication reconciliation, adverse reactions, non-adherence, or challenges related to the patient’s pharmacological regimen.
3. Condition complications/Disease progression: worsening or instability of a patient’s primary diagnosis (like Congestive Heart Failure (CHF), Chronic Obstructive Pulmonary Disease (COPD), or sepsis) shortly after the discharge.
4. Social determinants of Health (SDOH): lack of transportation to follow-up appointments, food insecurity, housing instability, and low income/insurance coverage…
5. Post-surgical complications/Falls: leading to new injuries, complications, and prolonged hospital stays.

Based on the top 5 hospital readmission factors, the following are the prioritized algorithms and the reasons for their prioritization:

Target Issue	Algorithm Types	Priority Algorithms
Non-compliance/lack of follow up	Prediction/Classification	XGBoost, Random Forest, CatBoost
Why these algorithms? These ensemble methods (tree-based) excel at handling complex, non-linear relationships in data, which is common when combining clinical, social, and demographic factors that predict patient behavior. They generally offer high predictive accuracy.

Target Issue	Algorithm Types	Priority Algorithms
Medication issues	Prediction/Classification & Explainable AI (XAI)	Logistic Regression (for interpretability) SHAP/LIME (XAI tools)
Why these algorithms? Logistic Regression is simple and highly interpretable, making it easy for pharmacists and doctors to understand why a medication risk is flagged (e.g., due to polypharmacy, specific drug interactions). XAI tools like SHAP/LIME are crucial to make complex models’ predictions transparent.

Target Issue	Algorithm Types	Priority Algorithms
Condition complications/Disease progression	Prediction/Classification	Neural Networks
Why these algorithms? For complex, time-series data like lab results, vital signs, and EHR notes, Deep Learning models (like Recurrent Neural Networks or Transformers) can capture subtle, evolving patterns of deterioration that simpler models might miss, offering superior prediction of clinical decline.

Target Issue	Algorithm Types	Priority Algorithms
Social determinants of Health (SDOH) and Demography:	Risk Stratification	Clustering Algorithms (e.g., K-Means, DBSCAN)
Why these algorithms? These are used for unsupervised learning, grouping patients into distinct, high-risk subpopulations based on factors like geographic location, cross-referencing local climate (e.g. transportation), local diet (e.g. food to avoid), insurance, and social needs. This allows for customized, non-clinical interventions (e.g., social work, transportation assistance).

Target Issue	Algorithm Types	Priority Algorithms
Post-surgical complications/Falls	Prediction/Classification & Explainable AI (XAI)	Random Forest, Decision Trees, SHAP/LIME
Why these algorithms? Decision Trees and Random Forests provide insights into the hierarchy of risk factors (e.g., which combination of age, surgery type, and frailty is most predictive of a fall or wound infection). The resulting model can be used to generate clear, actionable discharge instructions.

Though it is possible that models can be shared for various purposes, modular development is key to stitching functions together to complete the landscape and ensure future scalability. On the other hand, the team might also want to look into datasets, which will also be a crucial factor in model prioritization. For example, while looking into models for medication issues, one needs to further prioritize features and data points. The following are some typical feature candidates to consider:

1. Pharmacological: Total number of unique medications (Polypharmacy); total number of medication changes during admission; high-risk medications (e.g., opioids, anticoagulants, insulin).
2. Socioeconomic: Insurance status (proxy for medication cost/access), Patient zip code/Socioeconomic Status (SES) (a proxy for health literacy/support system).
3. Behavioral: History of non-adherence (if documented in prior clinical notes); number of unfilled prescriptions after previous discharges.
4. Clinical: Renal/Hepatic lab values at discharge, diagnosis (e.g., psychiatric conditions, diabetes, CHF) which are known to have high medication non-adherence rates.

Finally, to ensure leadership understands both the benefits and limitations of the recommended system, there are three practical categories to consider:

• Strategic value and ROI are the top concerns senior management will address. This includes:
  • Quantifiable financial benefit (ROI)
  • Enhanced clinical outcomes and patient experience
• Operational and technical feasibility: ensuring the AI system can be built and integrated without disrupting current operations. This includes:
  • Data readiness
  • System integration
  • Human resources
  • Phased approach deployment
• Risk & ethical governance: regulatory, ethical, and safety concerns are high-stakes in AI construction. Compliance is essential for the project to be successful for the leadership board. Compliance measures need to be set up for all phases thoroughly to minimize risk throughout the system development process. This includes:
  • Patient safety and accountability
  • Regulatory compliance
  • Bias & Fairness
  • Transparency & Trust

Once past the strategy level of reporting, for a mid-sized hospital, the following building blocks are recommended:

1. In the preparation stage, form a task force consisting of business development, marketing, project manager, subject experts (clinician, practitioner, nurse, lab…), IT (architect for system landscape, experience designer, developer, dbA, test…)
2. Define roles & responsibilities (R&R)
3. Do a series of Design Thinking sessions and define “problem to be solved” based on the top 5 areas mentioned above
4. Set product principles (project boundary, in-scope and out-of-scope activities, etc.)
5. Define target users and sample data (a limited sample size of 100 is sufficient for a PoC)
6. Try plotting a small-scale PoC to start with
7. Update leadership board (making sure they are in the loop and involved)
8. Evaluate both internal & external available datasets
9. Coding, test, and reiterate
10. Report the PoC results to leadership board and showing high confidence of successful execution for possible full-scale development support and funding.

Conclusion

In conclusion, leveraging a modular AI framework to target the five primary causes of readmission is not merely a technical exercise but a strategic imperative for improving patient care and achieving substantial financial returns in the current U.S. healthcare climate. By aligning algorithms like XGBoost and XAI tools with specific issues—from non-compliance to SDOH—the proposed system offers both high predictive power, clinical interpretability, and transparency. Success requires securing leadership commitment by rigorously addressing ROI, operational feasibility, and ethical governance. The recommended phased deployment, beginning with a structured Proof-of-Concept, is the necessary roadmap to transition this essential strategy into actionable, life-saving reality.

References

[1] Jasninder S. Dhaliwal; Ashujot Kaur Dang. (2024). Reducing Hospital Readmissions, https://www.ncbi.nlm.nih.gov/books/NBK606114/#:~:text=Lack%20of%20patient%20education:%20Many,plans%20and%20follow%2Dup%20care.