Predictive Risk Stratification for Chronic Disease Management
Clínica Las Condes had no systematic way to identify which Type 2 diabetes patients were at highest cardiovascular risk. I designed the clinical criteria, built the predictive model from paper records, and architected the full process, achieving 94.95% classification accuracy on 1,487 patients.
The problem
Private healthcare in Chile is a low-differentiation market. Clinics compete on price and infrastructure, not on patient outcomes. Clínica Las Condes had a strategic ambition to shift from reactive, episode-based care to a proactive, outcome-driven model for patients with chronic conditions.
The specific gap: Type 2 diabetes patients were not being stratified by cardiovascular risk. Every patient received the same clinical attention regardless of actual risk level. High-risk patients weren't getting early intervention. Clinical resources were distributed inefficiently across all risk groups equally. And the clinic had no systematic way to identify which patients would likely generate the highest cost burden and complications.
The Director of Medical Management knew this needed to change. The question was how to build a scalable, clinician-usable risk classification system when patient data lived entirely in paper clinical records. Nobody had defined the problem clearly enough to act on it. That was my starting point.
My scope
This project was not handed to me as a defined brief. I was embedded with the Directorate of Medical Management and owned the full arc:
- Defining the problem: framing the strategic opportunity, the business case, and the scope of a viable solution
- Designing the end-to-end clinical process: from patient data entry through automated risk classification to differentiated clinical intervention
- Building the predictive model: data extraction from paper records, cleaning, preprocessing, and decision tree construction
- Designing the IT architecture: use cases, sequence diagrams, class diagrams, and a 3-tier web application blueprint
- Developing a generalizable framework: abstracting the solution into a reusable pattern applicable to any predictive risk model based on decision trees
- Designing the change management plan: addressing clinical staff resistance to moving from paper to digital workflows
Discovery: defining what "risk" actually means
The first and most critical decision was not technical. It was clinical and strategic: what does cardiovascular risk mean for a Type 2 diabetes patient at this institution, and how do we operationalize it in a way clinicians will trust and use?
I ran working sessions with the clinic's endocrinology specialists over several months. We reviewed the 2009 MINSAL Clinical Guide for Type 2 Diabetes, cross-referenced international evidence on cardiovascular risk factors in diabetic populations, and progressively built consensus around what the clinic's own "cut-off criteria" would be.
The output was a co-designed scoring system, "CLC Cardiovascular Risk Cut-Off Criteria for Type 2 Diabetic Patients," that translated clinical judgment into a computable rule set, segmenting patients into three risk groups: low, medium, and high. The final variable set included: HbA1c, HDL cholesterol, total cholesterol, triglycerides, hypertension status, prior coronary disease, smoking, age, and sex.
This was not a standard scoring tool lifted from literature. It was co-created with the clinicians who would use it. By grounding the classification logic in expert consensus rather than purely statistical inference, the resulting decision tree was interpretable: clinicians could read why a patient was classified at a given risk level, not just what the output was. Clinical adoption requires explainability, not just accuracy. That was a deliberate design decision.
The data challenge
The raw data came from paper clinical records and physical lab results for patients seen by diabetologists at Clínica Las Condes during 2008. This created significant data quality challenges before any modeling could happen: missing values across clinical variables, outliers from transcription errors (including biologically impossible lab values), and no standardized input format.
After cleaning, preprocessing, and binarizing continuous variables into clinically meaningful thresholds, the final dataset comprised 1,487 labeled patient records: 474 high-risk, 531 medium-risk, 482 low-risk, reasonably balanced across classes.
Data cleaning was one of the most resource-intensive phases of the project, a constraint I documented explicitly, because it revealed a critical systemic gap: without structured digital data capture, predictive modeling at scale would remain impossible. That finding shaped the IT architecture recommendations.
Solution: a process-integrated classification system
The model alone was not the deliverable. The deliverable was a redesigned clinical process with the model embedded in it. Using BPMN process modeling and business engineering methodology, I designed three interconnected macro-processes:
The decision tree was built using the J48 algorithm in Weka Workbench, with stratified cross-validation to account for the limited dataset size and avoid overfitting. The supporting IT architecture used a 3-tier web application model with two user roles: Clinician (data entry and patient evaluation) and Medical Director (performance indicators and model parameter review).
Results
Model performance, stratified cross-validation, 1,487 patients:
| Metric | Value | Note |
|---|---|---|
| Overall accuracy | 94.95% | 1,412 of 1,487 correctly classified · Tabla 13, p. 171 |
| Incorrectly classified | 75 records (5.04%) | Stratified cross-validation |
| Risk Class | Precision | Recall | F-Measure |
|---|---|---|---|
| High risk | 0.972 | 0.943 | 0.958 |
| Medium risk | 0.951 | 0.950 | 0.951 |
| Low risk | 0.924 | 0.955 | 0.939 |
The high-risk class achieved the strongest precision (0.972). This matters clinically: a false negative in a high-risk patient, missing someone who will go on to have a cardiac event, carries a significantly higher cost than a false positive. The model's asymmetric error performance was evaluated explicitly against clinical consequence. That's a guardrail metric, not an afterthought.
The generalizable framework
Beyond the clinical result, I abstracted the full methodology into a reusable Decision Tree Risk Framework, a pattern applicable to any organization wanting to build predictive risk classification using decision trees. This included class diagrams, business logic specifications, and implementation guidelines adaptable to other chronic conditions or other institutions.
What didn't happen, and why it matters
The system was not implemented in production at Clínica Las Condes.
The barrier was not technical or clinical. It was structural. The clinic's existing care model was purely reactive and episode-based. Adopting a proactive disease management approach would have required a fundamental shift in how the institution understood its business model, not just a new IT system. At the time, no Chilean private healthcare provider had institutionalized this kind of preventive care program.
The change management plan I designed (covering coalition building, narrative strategy, communication rollout, and success criteria) addressed this. But without budget allocation and executive sponsorship as entry conditions, the proof of concept remained a proof of concept.
The most important lesson from this project: the hardest constraint in clinical product work is rarely the model. It's organizational readiness: the gap between what the evidence supports and what the institution is structured to adopt. Every subsequent project I've led in health systems transformation has been shaped by this.
What I'd do differently
Start with the data infrastructure problem, not the model. The dependency on paper records was the actual bottleneck. A phased approach, first digitizing intake, then modeling, would have created faster institutional buy-in and a cleaner dataset.
Define the implementation owner earlier. The project had a clear technical and clinical owner but no operational owner inside the organization committed to running the program post-handoff. That gap is a product failure, not a research limitation.
Propose a smaller pilot scope. Starting with a single physician cohort rather than an institution-wide redesign would have reduced the change management threshold and created early proof points that could build momentum.
This project was developed as part of the MBE at Universidad de Chile and cited in academic publications co-authored with Prof. Óscar Barros (Serie Sistemas Públicos N°14). Academic committee: Prof. Óscar Barros, Dr. Emilio Santelices, and two additional faculty reviewers.