TeachAhead

TeachAhead: Predicting Change, Preserving Identity

Inspiration

Amyotrophic lateral sclerosis (ALS) is a progressive neurodegenerative disease that gradually affects essential functions such as speaking, writing, swallowing, and walking. Although clinical assessments can measure functional decline, these measurements do not fully capture how the disease affects a person's identity, relationships, and everyday goals.

We wanted to explore how machine learning could help anticipate functional changes while keeping the individual behind the clinical data at the centre of care.

Our project was inspired by Alex Chen, a fictional 47-year-old biology teacher living with ALS. Alex experiences increasing difficulties with speech, handwriting, and mobility, but his priorities extend beyond clinical measurements. He wants to continue teaching, prepare Sunday meals for his family, and remain actively involved as a father and partner.

This led us to an important question:

What if predicting disease progression was only the first step, and we could also help patients and healthcare providers plan around what matters most to each individual?

That question became the foundation of TeachAhead.

What It Does

TeachAhead is a research prototype combining machine learning-based ALS progression insights with an interactive, patient-centred planning dashboard.

Our project consists of two complementary components.

ALS Progression Prediction and Verification

We developed a machine learning model using the competition's ALS Dataset 2, containing 1,822 patient records.

The model uses 22 early clinical features to estimate an individual patient's average rate of functional change between months 3 and 12 after trial entry.

These features include measurements related to functional status, speech, swallowing, mobility, respiratory function, and weight.

We also developed a separate verification interface to examine historical predictions, observed outcomes, prediction errors, and uncertainty.

Our selected Clinical Random Forest model achieved the following results on 625 historical evaluation patients:

RMSE: 0.5093 points/month

MAE: 0.3991 points/month

RMSE improvement: Approximately 10% compared with a constant-mean baseline

Empirical 90% prediction interval coverage: 88.96%

Notably, the 22-feature model achieved performance comparable to a Random Forest model using 369 candidate predictors.

These findings suggest that a smaller set of early clinical measurements may retain useful predictive information, although further validation is necessary.

Patient-Centred Planning Dashboard

We built an interactive Streamlit dashboard to demonstrate how clinical monitoring, personal goals, and patient feedback could support individualized care planning.

Using Alex's fictional case, the dashboard includes four main sections:

ALS Progression: Visualizes Alex's observed ALSFRS-R history alongside separately evaluated historical machine learning results. Users can explore predictions and uncertainty intervals for historical dataset records.

Current Functional Status: Presents Alex's functional challenges, including speech, handwriting, mobility, fatigue, and respiratory status, while highlighting his personal goal of continuing to teach.

Patient-Centred Planning: Provides rule-based support options based on the patient's main challenges and preferences. For example, if Alex experiences speech difficulties, the dashboard suggests discussing classroom voice amplification, communication strategies, and speech-language support.

Feedback & Follow-Up: Allows users to record new clinical assessments, patient experiences, and caregiver observations. The observed progression chart updates when new functional scores are entered, while the Plan Review section provides prompts to reconsider strategies that patients find uncomfortable or unhelpful.

For example, if Alex tries a classroom microphone but finds it uncomfortable, the dashboard encourages discussion of alternative strategies rather than assuming the original recommendation should continue.

Our goal is to demonstrate that effective care planning should respond not only to changes in clinical measurements but also to the patient's lived experience.

How We Built It

Our project combines machine learning analysis in Python with an interactive web application developed using Streamlit.

Machine Learning Development

We used ALS Dataset 2, containing 1,822 records and 369 candidate predictors, along with the functional progression target.

The data were divided into:

957 patients for model development

240 patients for calibration

625 patients for historical evaluation

We compared several machine learning approaches, including Ridge Regression, Support Vector Regression, Random Forest, Histogram Gradient Boosting, and Decision Tree models.

After model comparison and cross-validation, we selected a Clinical Random Forest using 22 early clinical features, with 200 trees, a maximum depth of 6, and a minimum leaf size of 15.

We used split conformal calibration to construct prediction intervals, allowing us to communicate uncertainty rather than presenting predictions as exact outcomes.

We also evaluated the model against a constant-mean baseline and a larger 369-feature Random Forest.

Dashboard Development

We developed the TeachAhead interface using:

Python for application logic and data processing

Streamlit for the interactive web application

Pandas for organizing and displaying clinical and prediction data

Altair for visualizing functional changes and historical model predictions

GitHub and Streamlit Community Cloud for sharing the application

We designed the dashboard around four principles:

Predict → Personalize → Monitor → Adapt

The prediction section displays saved results from our evaluated machine learning model. The personalized planning section uses predefined rules based on Alex's challenges and preferences.

The monitoring and feedback components use Streamlit session state to update the displayed clinical history and plan review prompts during a demonstration.

Importantly, the dashboard does not retrain the machine learning model when new clinical measurements or patient feedback are entered.

Challenges We Faced

One of our biggest challenges was connecting machine learning results with meaningful patient-centred applications.

Although the model could estimate an individual's average functional progression rate, this did not directly tell us how to support specific everyday activities, such as teaching, communicating with family, or maintaining independence.

We also faced an important data compatibility challenge. Our machine learning model predicts signed dFRS on the competition's original 0–40 functional scale, while Alex's fictional case includes ALSFRS-R measurements on a 0–48 scale.

Because these measurements are not directly interchangeable, and Alex's case does not contain all the required early clinical inputs, we deliberately avoided generating an unsupported machine learning prediction for him.

Instead, we clearly separated historical model predictions from Alex's observed clinical history.

Another challenge was translating complex clinical information into an accessible, interactive dashboard. We wanted users to understand functional changes without losing sight of the patient's preferences and goals.

Finally, we had to ensure that our planning features were presented responsibly. The suggested support strategies are rule-based discussion options, not automated medical recommendations.

What We Learned

This project helped us understand that predictive performance is only one part of developing a useful healthcare technology.

We learned how to compare machine learning models, evaluate prediction errors, communicate uncertainty, and recognize the importance of independent validation.

We also learned that clinical data alone cannot capture what matters most to a patient.

Two people with similar functional limitations may have very different priorities, preferences, and support needs.

Through developing TeachAhead, we explored how patient feedback and repeated clinical assessments could complement predictive research and support more responsive care-planning discussions.

Most importantly, we learned that a patient-centred approach requires listening to individuals rather than relying exclusively on numerical measurements.

What's Next?

TeachAhead is currently a research and demonstration prototype, not a clinically validated decision-support tool.

Our next steps would include:

Integrating patient-level predictions with the planning interface: Develop a compatible data pipeline that can connect model outputs to individual patient assessments without assuming equivalence between different functional scales.

Independent model validation: Evaluate predictive performance on external ALS cohorts and assess generalizability.

Patient and clinician feedback: Conduct usability testing with patients, caregivers, and healthcare professionals to improve the dashboard's accessibility and relevance.

Longitudinal monitoring: Explore how repeated assessments and patient-reported experiences could support ongoing care-planning discussions.

Clinical implementation considerations: Investigate privacy, security, clinical oversight, and regulatory requirements before any real-world deployment.

Our long-term vision is to explore how predictive information and patient preferences can work together to support proactive, individualized ALS care.

TeachAhead is not just about predicting change. It is about helping people preserve the activities, relationships, and identities that matter most to them.