AIdeas Finalist: Predict-Epidem
Predict-Epidem is an epidemiological intelligence system designed to monitor and predict outbreaks in Latin America. Through a high-fidelity dashboard, it provides real-time data visualization, risk prediction, and critical alerts that transform crisis management into a precision strategy. By integrating climate data, clinical reports, and social media signals, it identifies contagion patterns, allowing healthcare systems to act before medical units reach their limit.
App Category
Social Impact
Our Vision
Predict-Epidem: Weeks of advantage, lives of difference
Latin America is experiencing an unprecedented epidemiological crisis. In 2023 alone, Mexico reported more than 54,000 cases of dengue, a 230% increase compared to previous years, and the upward trend continued in 2024 with the resurgence of the DENV-3 serotype, which had not predominated for years, causing massive hospitalizations in critical states such as Veracruz (where 4 serotypes circulate) and Guerrero. This scenario reveals a structural flaw: the major gap in public health is not the lack of data, but the inability to act in time. Currently, hospitals operate under a reactive model, only mobilizing when the wards are already full and contagion is exponential. As the literature in The Lancet Digital Health points out, the real obstacle is not Artificial Intelligence, but rather that our systems were designed to archive past records and not to feed models that predict the future in real time.
Predict-Epidem was created to break this inertia, transforming crisis management into a precision strategy through a deep learning architecture on AWS. By analyzing four decades of epidemiological data from SINAVE (Sistema Nacional de Vigilancia Epidemiológica de México, 1984-2025), our model cross-references this information with climate patterns, biological cycles, and, crucially, social media signals. These signals act as an immediate response sensor, detecting mentions of symptoms and outbreaks in specific communities long before patients arrive at an official clinic.
This multivariate integration allows us to identify early signals invisible to traditional analyses, projecting outbreaks with a window of 2 to 4 weeks' notice. Our vision begins in Mexico, but it aims at the entire region: we want to ensure that no vulnerable community in Latin America is ever again taken by surprise by a predictable epidemic.
To address data fragmentation, we implement a multi-source triangulation strategy:
epidemiological records are cross-validated against climate anomalies and anonymized social media volume trends. The WHO (World Health Organization) estimates a dengue underreporting factor of 3-5x (Brady et al., 2012); our model accounts for this by training on relative trends rather than absolute case counts, making predictions robust to systematic underreporting.
Regarding social media signals, we emphasize that Predict-Epidem processes only aggregate, anonymized keyword frequency data, no individual posts, usernames, or personally identifiable information (PII) are stored or analyzed. This approach follows established methodologies from HealthMap (Boston Children's Hospital) and ProMED-mail, which have demonstrated the value of digital surveillance as a complementary data source for over a decade (Brownstein et al., 2009)
Why this matters
Mexico and Latin America face a cyclical threat from arboviruses such as dengue, Zika, chikungunya, and the emerging Oropouche virus. In regions with high climate vulnerability, these outbreaks drive systemic hospital collapse: saturated wards force the postponement of scheduled surgeries, overwhelm obstetrics units, and deplete critical supplies for unrelated emergencies
Predict-Epidem breaks this cycle through a dual-temporal approach that responds to the real needs of public administration. First, it enables annual strategic budget planning by identifying high-risk years based on long-term climate cycles, facilitating year-over-year resource allocation. Second, it triggers an operational alert two to four weeks in advance, the critical window needed for institutions to implement preventative measures before an outbreak reaches its peak.
By anticipating risk weeks ahead, our solution reduces hospital overcrowding and minimizes costly emergency responses, closing preparedness gaps that have historically cost lives. Scientific evidence supports this: a systematic review of 48 early warning systems across 32 countries found that EWS strengthened healthcare resilience by 15-25%, with 2-4 week advance warnings enabling targeted vector control and hospital pre-positioning (Meckawy et al., 2022, BMC Public Health)
Why is the Predict-Epidem different?
This solution goes beyond data analysis; it leverages the scalability of cloud technology to build a pillar of resilience at the heart of the healthcare system. By processing and converting the "noise" of fragmented data, weather variables, and real-time social media signals into high-fidelity predictive intelligence, Predict-Epidem ensures the continuity of essential medical services. Our proposal guarantees that hospital infrastructure not only survives epidemiological peaks but also maintains its full operational capacity, protecting both outbreak patients and those who depend on uninterrupted critical care. Ultimately, we transform data-driven prevention into a safeguard for regional health stability.
To understand Predict-Epidem's value proposition, it's important to examine the current landscape of epidemiological surveillance tools, The key insight is that existing tools were designed to archive and report historical data, not to predict future outbreaks.
Feature | SINAVE | WHO DengueNet | Google Dengue Trends | HealthMap | Predict-Epidem |
|---|---|---|---|---|---|
System Type | Reactive | Reactive | Discontinued | Reactive | Predictive |
Prediction Window | None | None | N/A | None | 2-4 weeks |
Climate Data Integration | No | No | No | Partial | Yes |
Automated Alerts (SMS) | No | No | No | Email Only | Yes (SNS) |
Cost | Gov.Only | Institucional | N/A | Academic | $0 Free Tier |
State-level Granularity (MX) | Yes | No | No | No | Yes |
Open Source | No | No | No | Partial | Yes (MIT) |
How We Built This
We implemented a 100% serverless, event-driven architecture on AWS, designed to scale instantly during outbreak peaks without the operational burden of managing servers.
The Tech Stack
Layer | Tool | Critical Function |
|---|---|---|
Core & AI IDE | Kiro AI IDE | Development orchestration, data pre-processing assistance, and medical context management |
Machine Learning | Amazon SageMaker Canvas | Time-series prediction with high-fidelity algorithms and automatic training. |
Compute & Logic | AWS Lambda | Serverless processing, data normalization, and alert trigger execution. |
Data Lake | Amazon S3 | Central repository for historical epidemiological and climate data (2020–2025). |
Automation | Amazon EventBridge | Daily automation of climate data ingestion and event coordination. |
Visualization | Amazon QuickSight | Interactive geographic risk dashboards and BI reports for decision-making. |
Communication | Amazon SNS | Automated SMS alerts to medical personnel in areas with low connectivity. |
Critical Architecture Decisions Why not use a relational database (RDS)?
- S3 + Parquet offers better cost-efficiency for time series analysis.
- Compliance with Free Tier (RDS consumes credits quickly).
Why SageMaker Canvas over Custom Models?
- Time-to-market: Canvas reduces development by 3 weeks.
- Pre-optimized algorithms (DeepAR+ & Prophet) with WAPE <0.15. No-code allows scalability without dedicated DevOps.
Methodology: Spec-Driven Development with Kiro AI
To accelerate development and ensure technical consistency, we used the Kiro AI IDE. This approach allowed us to prioritize architecture over code:
- Spec Definition: We structured the project using detailed specifications (Data Pipeline, ML Pipeline, and Alert System). Kiro transformed these business rules into robust AWS Lambda functions with integrated error handling and preprocessing logic.
- Continuous Epidemiological Context: By using Steering Files, we ensured that Kiro maintained context regarding dengue's biological cycles and climate seasonality. This was crucial for the AI to generate the data cleaning code (report parsing and normalization) necessary to feed the machine learning model.
Predictive Intelligence: Precision over Intuition
During the design phase, we selected Amazon SageMaker Canvas as the project's predictive core, discarding traditional architectures in favor of a high-speed, visual machine learning approach. Given the maturity of AWS solutions in 2026, Canvas emerged as the logical choice due to its ability to efficiently process complex time series.
- High-Precision Modeling: The system automatically evaluated and selected the most robust algorithms (DeepAR+ and Prophet), integrating 40 years of SINAVE data and climate variables in less than 3 hours of training.
- Technical Validation: We obtained a WAPE of 0.12 and a MAPE of 0.15, metrics that validate our ability to project epidemiological peaks 4 weeks in advance, allowing the results to be visible in real time using Amazon QuickSight.
Architecture Diagram
Validation & Backtesting
To validate our model’s predictive capability beyond standard ML metrics, we conducted historical backtesting using a walk forward validation approach.
Walk-Forward Validation Methodology
Unlike traditional cross-validation, which randomly splits data and can introduce future data leakage in time series, walk-forward validation preserves temporal order and simulates how the model would have performed in a real deployment scenario:
- Training window: The model trains on all available data up to week N, starting with a minimum of 52 weeks (1 full year) of historical data
- Prediction: The model forecasts disease case counts for the next 4 weeks (N+1 to N+4)
- Evaluation: Predictions are compared against actual reported cases from SINAVE
- Advance: The window shifts forward by 4 weeks, and the process repeats
- Outbreak detection: A Z-score threshold of 1.5 flags periods where predicted case counts significantly exceed historical baselines
This process was executed across 52 iterations for dengue (using 5 years of real epidemiological data from SINAVE combined with climate variables from Veracruz) and 442 iterations for COVID-19 (using real case data from 2020–2024).
Historical Backtesting: The 2023 Dengue Surge
The most compelling validation came from the 2023 dengue epidemic in Mexico, which saw a 230% increase in cases compared to 2022. Our model, trained only on data available before each prediction window, successfully detected the escalating outbreak pattern 4 weeks in advance at each stage:
Training Data Cutoff | Predicted Cases (4-week avg) | Actual Cases (4-week avg) | MAE | Detection |
|---|---|---|---|---|
April 23, 2023 | 1,262 | 1,299 | 44 | Detected |
May 21, 2023 | 2,183 | 2,354 | 171 | Detected |
June 18, 2023 | 3,809 | 3,839 | 176 | Detected |
July 16, 2023 | 5,749 | 6,389 | 640 | Detected |
August 13, 2023 | 13,104 | 13,321 | 2,546 | Detected |
The model captured the full trajectory of the outbreak, from early escalation in April through the peak in August, with a mean absolute error as low as 44 cases during the initial detection phase. Every outbreak period in the 2023 surge was correctly flagged, providing the type of advance warning that enables healthcare systems to pre-position resources, activate vector control programs, and prepare hospital capacity before cases overwhelm the system.
Overall Outbreak Detection Performance
Across all 52 walk-forward iterations on real dengue data, the model achieved:
Metric | Value | Interpretation |
|---|---|---|
Precision | 73.3% | When the model flags an outbreak, it is correct ~3 out of 4 times |
Recall | 78.6% | The model detects ~4 out of every 5 actual outbreaks |
F1 Score | 0.759 | Strong balance between false alarms and missed outbreaks |
True Positives | 11 | Correctly detected outbreaks |
False Positives | 4 | False alarms (model predicted outbreak that didn't meet threshold) |
False Negatives | 3 | Missed outbreaks |
For COVID-19, the model achieved a MAPE of 9.6% across 442 walk-forward iterations using real case data demonstrating that the architecture generalizes across different disease dynamics, not just vector-borne illnesses.
Transparency note: Backtesting for influenza, chikungunya, and zika used synthetic datasets generated from epidemiological distribution patterns for proof-of-concept purposes. We report only real-data validation metrics here. Integration of real historical data for these diseases from SINAVE and PAHO is prioritized in our Phase 2 roadmap.
While historical backtesting demonstrates retrospective accuracy, we recognize that real-world deployment validation is essential. Our field validation strategy includes:
- Public Health Advisory: We have engaged a former Director of the Guanajuato State Health Ministry as a public health advisor to guide our validation methodology and provide institutional credibility. Their expertise in epidemiological surveillance at the state level ensures our approach aligns with Mexico's existing public health infrastructure and reporting protocols.
- Institutional Validation Letter: We are working with our public health advisor to produce a formal expert validation of our methodology and a letter of intent for pilot collaboration with state health authorities.
- Publication: Pilot results will be submitted as a peer-reviewed case study to a public health journal (target: BMC Public Health or PLOS ONE), providing independent scientific validation of the platform's predictive capabilities.
DEMO
What We Have Learned
The development of Predict-Epidem has been a journey of bridging the gap between cloud engineering and epidemiological intelligence. Our key lessons center on how a serverless architecture can transform raw public health data into actionable life-saving windows.
Technical Lessons and Data Engineering
Kiro as a Data Architect: We learned that the value of Kiro Steering Files was crucial in the pre-processing phase. Allowed us to code epidemiological business rules, such as calculating moving windows, handling seasonality, and normalizing incidence rates, during the creation of the training and validation datasets. Without this context injected into the Lambda functions, the raw data would have lacked the necessary structure for SageMaker Canvas to detect significant geographic patterns.
SageMaker Canvas Effectiveness in 2026: We confirmed that for high-fidelity prototyping, SageMaker Canvas completed training on 40 years of historical epidemiological data approximately 2,100 weekly records spanning five diseases in 2.8 hours, measured during our SageMaker Canvas training session. By comparison, equivalent custom model development (data preprocessing, feature engineering, hyperparameter tuning, and validation) on a comparable ML instance typically requires 15–20 hours of engineering time with a custom approach in EC2, saving approximately $45 in compute costs.
By delegating algorithm selection, we were able to focus our efforts on feature engineering (identifying that relative humidity and minimum temperature are more powerful predictors than rainfall itself).
Cloud Viability: We validated that a 100% serverless architecture allows us to operate a nationally accessible MVP under the AWS Free Tier, eliminating the economic barrier to innovation in public health.
Epidemiological Domain and Real-World Challenges
The Critical Lead Time: We confirmed that climate variables offer a prediction window of 4 to 8 weeks. Our walk-forward backtesting validated a 4-week forecast horizon with an outbreak detection recall of 78.6% and precision of 73.3%. The extended 8-week window is supported by climate-lag analysis showing that temperature and precipitation anomalies correlate with dengue case surges at 6–8 week offsets (Lowe et al., 2018, The Lancet Planetary Health). This learning is vital, as a hospital requires a minimum margin of 14 days to reconfigure its logistics of beds and critical supplies.
Data Ingestion Challenges: The biggest technical obstacle is the nature of the public data; many historical SINAVE reports still reside in unstructured formats (PDF), requiring robust parsing processes to avoid bias in the model. We also learned to manage the limitations of mass communication, optimizing the use of Amazon SNS to send alerts only when risk thresholds exceed 80%.
Responding to Judge Feedback
The semifinal evaluation was a turning point for Predict-Epidem. While judges validated our technical foundation and social impact vision, their suggestions for refinement revealed critical gaps between a promising prototype and a credible public health tool. Here is how we addressed each one:
"Limited validation evidence" This resonated the most with our team. We realized that standard ML metrics alone (WAPE 0.12, MAPE 0.15) were insufficient to demonstrate real-world readiness. In direct response, we designed and executed a walk-forward backtesting study using real SINAVE dengue data from 2020–2024. The results were decisive: our model successfully detected every stage of the 2023 dengue surge, the worst in Mexico's recent history, with 4 weeks of advance warning and an outbreak detection recall of 78.6% (see Validation & Backtesting section above). We also engaged a former Director of the Guanajuato State Health Ministry as a public health advisor, bringing institutional credibility and a clear path toward field validation with state health authorities.
"Competitive landscape not addressed" We conducted a systematic comparison against five existing epidemiological tools (SINAVE, WHO DengueNet, Google Dengue Trends, HealthMap) and discovered that every current solution operates reactively, archiving past data rather than predicting future outbreaks. This analysis sharpened our value proposition: Predict-Epidem is the only tool that combines predictive modeling, climate integration, automated SMS alerts to low-connectivity zones, and zero-cost deployment, all within the AWS Free Tier (see Competitive Landscape table in Why This Matters).
"Social media privacy and data quality concerns" Judges raised valid questions about how we handle fragmented data and social media signals. We strengthened our approach by implementing a multi-source triangulation strategy: epidemiological records are cross-validated against climate anomalies and anonymized social media volume trends. Critically, we clarified that our platform processes only aggregate keyword frequency data, no individual posts, usernames, or PII are stored at any stage, following established precedents from HealthMap and ProMED-mail (Brownstein et al., 2009, NEJM). For data quality, we addressed the WHO's estimated 3–5× underreporting factor by training on relative trends rather than absolute counts, making predictions robust to systematic underreporting (Brady et al., 2012, Nature)
"Technical claims lack supporting evidence" Every quantitative claim in our revised article now includes its source: the 15–25% healthcare impact cites a systematic review of 48 EWS across 32 countries (Meckawy et al., 2022); the prediction window is validated by our own backtesting; and the 2.8-hour training time was measured from our actual SageMaker Canvas session processing approximately 2,100 weekly records across five diseases.
These refinements transformed Predict-Epidem from a compelling idea into an evidence-backed epidemiological intelligence platform, one where every claim is substantiated, every limitation is acknowledged, and every design decision is driven by the operational reality of Latin America's healthcare systems.
Roadmap: Next 6 –12 Months
Phase 1 (Q2 2026): Field Validation
- Controlled pilot study with 3 hospitals in high-incidence municipalities of Veracruz (Xalapa, Coatzacoalcos, and Poza Rica), selected based on historical case volume (top 15% state-wide) and existing digital reporting infrastructure.
- Integration with local HIS (Hospital Information System) for real-time case count validation against model predictions, running in shadow mode (predictions generated but not yet used for clinical decisions) for the first 8 weeks.
- Engagement with state health authorities through our public health advisor a former Director of the Guanajuato State Health Ministry to ensure regulatory alignment and institutional buy-in.
- Feedback loop: Hospital staff input on alert format, timing, and actionability to refine the notification system before Phase 2 scale-up.
Phase 2 (Q3 2026): Signal Expansion
- Google Trends API for early detection
- Integration with X for self-reported symptom analysis
Phase 3 (Q4 2026): Multi-Pathology
- Extension to Zika, Chikungunya, and Oropouche
- Unified model of arboviruses transmitted by Aedes aegypti
Phase 4 (2027): Regional Scale-Up
- Expansion to Colombia, Chile, and Argentina
- Collaboration with OPS/OMS for data standardization
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article