Personalized Allergy Prediction

01 • Research

The question

Generic pollen forecasts assume people react similarly. This project tests whether a model trained on one person's own history can make a more useful next-day prediction.

Can a personalized machine-learning model that combines pollen, weather, air quality, indoor dust data, and daily symptom history reliably predict next-day allergy symptom severity and provide an early warning?

Hypothesis. A neural network using personal history plus environmental and indoor data will predict next-day symptom severity more accurately than a pollen-only warning.

76input features across environment, habits, and sensors

11symptom outputs predicted by the model

221recent low-pollen days used in the comparison chart

Dailyautomated collection, retraining, prediction, and alerts

02 • Inputs

What the model sees

The project intentionally goes beyond pollen. Outdoor conditions, indoor air, behavior, and symptom reports are brought together into one daily dataset.

🌾

PollenTree, grass, weed, and individual pollen indices.

🌦️

WeatherTemperature, precipitation, wind, humidity and related variables.

🌫️

Air qualityAQI plus pollutants such as PM2.5, PM10, O₃ and NO₂.

🏠

Indoor particlesSPS30 particle measurements collected in the home.

💊

MedicationDaily medication and exposure context.

🥗

Food & habitsManual logs that capture potential non-pollen influences.

🤧

SymptomsSeverity reports such as itchy eyes, sneezing, fatigue, congestion and more.

📈

Trend signalsOther external indicators, including optional Google Trends data.

Table of project input variables from weather, pollen, air quality, SPS30 sensor, food and symptom tracking — Input-variable inventory used by the project.

03 • Engineering

How predictions happen

A scheduled pipeline turns the experiment into a repeatable daily system: collect → store → train → predict → show an alert.

Engineering process diagram for the allergy prediction system — End-to-end architecture: Apps Script, Google Sheets, TensorFlow/Keras, AppSheet, APIs, and Raspberry Pi sensor collection.

1

Data entry

Apps Script gathers pollen, weather, air quality, trend data, and indoor sensor readings, then standardizes the daily record.

2

Automation

Time-based triggers collect new data, validate the latest row, refresh calculations, retrain, and run predictions.

3

Database

Google Sheets stores the history of environmental variables, indoor readings, and personal symptom logs.

4

Machine learning

TensorFlow/Keras learns relationships between conditions and next-day symptom outcomes.

5

User interface

AppSheet handles daily logging and presents current predictions and high-risk warnings.

04 • Model

The neural network

The model converts a wide mix of inputs into separate symptom predictions. The architecture stays compact enough to retrain routinely as new personal data arrives.

76 input features40-neuron hidden layer10-neuron hidden layerReLUSigmoid output

Personalization is the key idea

Instead of asking only “Is pollen high?”, the network learns combinations that are specific to the user — for example wind + dryness + particles + past symptom patterns.

Evaluation

Training uses metrics including MAE, RMSE, and R². Predictions can also be converted to high-risk flags with a predefined cutoff.

Neural network model architecture with 76 input features, 40 and 10 neuron hidden layers, and 11 symptom outputs — Neural-network architecture used for next-day symptom prediction.

05 • Results

What the data showed

The training curves improved over time, while symptom-level evaluation showed that some outcomes were easier to forecast than others. The most practical comparison was against the simple pollen-only baseline.

Personalization beat the pollen-only warning on the low-pollen comparison.

Across the recent 221-day low-pollen subset shown in the project report, the personalized model made fewer incorrect predictions for itchy eyes, sneezing, fatigue, and yawning. That is the setting where a generic “low pollen = low risk” rule is especially likely to miss symptoms driven by other conditions.

This is a single-participant student study, so the result supports the project hypothesis but does not establish medical effectiveness for other people.

Bar chart comparing incorrect pollen-count predictions with incorrect personalized model predictions on 221 low-pollen days — Comparison of incorrect predictions in the low-pollen subset.

Training and validation loss decreasing over epochs — Training and validation loss both trend downward.

Training and validation mean absolute error decreasing over epochs — Mean Absolute Error decreases during training.

Symptom by symptom classification performance table — Symptom-level precision, recall, F1, accuracy, and confusion-count metrics.

F1 score ranking by symptom — F1 scores show that prediction performance differs by symptom.

06 • Presentation

Project walkthrough

A video presentation of the research question, data pipeline, model, and results.

07 • Conclusion

What it means

The project supports the idea that allergy forecasting can become more useful when it learns the way multiple environmental conditions interact for one person.

More than pollen.

The model combined pollen, weather, air-quality and particle measurements with personal symptom history. The report concludes that this multi-factor approach can identify symptom risk that a pollen-only warning may miss — particularly on low-to-moderate pollen days.

Key limitations: one participant, a limited collection window, self-reported symptoms, and missing variables such as exact outdoor exposure time, sleep, and detailed medication timing.

08 • Next steps

Future possibilities

The report groups future work into three directions: richer data, smarter modeling, and broader integrations.

More inputs

Add participants, collect across more seasons, expand sensor coverage, improve symptom labels, and incorporate medical guidance where appropriate.

Smarter modeling

Engineer stronger features, compare additional model types, deepen personalization, and strengthen privacy and security.

More integrations

Explore LLM-assisted explanations plus calendar, reminder, and collaboration integrations for predicted high-risk days.

09 • Sources

Selected references

Background research cited in the project report.

Alnahas, S. et al. (2023). Prevalence, severity, and risk factors of allergic rhinitis among schoolchildren in Saudi Arabia. World Allergy Organization Journal. DOI ↗
Ng, A. E., & Boersma, P. (2023). Diagnosed allergic conditions in adults: United States, 2021. CDC/NCHS. CDC ↗
Asthma and Allergy Foundation of America. (2025). Allergy Capitals. AAFA ↗
Osborne, N. J. et al. (2017). Pollen exposure and hospitalization due to asthma exacerbations. DOI ↗
Shrestha, S. K., Lambert, K. A., & Erbas, B. (2021). Ambient pollen concentrations and asthma hospitalization in children and adolescents. DOI ↗
Carlsen, H. K. et al. (2022). Birch pollen, air pollution and their interactive effects on airway symptoms. DOI ↗