week1.md

Week 1 packet for Amber: T2D phenotyping + EHR + neural networks

This packet is designed for a one-week bridge from applied math into biomedical informatics. The goal is not mastery in one week. The goal is to build a working mental model of:

1. what EHR / real-world data are,
2. what a computable phenotype is,
3. how T2D is defined clinically vs operationally in data,
4. why comorbidity and coding systems matter,
5. how a basic neural-network pipeline works in R.

Suggested effort: 10-12 hours total.

How to use AI productively during the rotation

Use ChatGPT Study Mode or any strong reasoning assistant you prefer as a coach, not as a replacement for reading.

Good uses:

• “Explain the PheKB T2D algorithm as a decision tree.”
• “Quiz me on diabetes diagnosis thresholds.”
• “Turn this notebook into a step-by-step checklist.”
• “Explain what overfitting would look like in my training curves.”
• “Give me a plain-English explanation of Charlson comorbidity.”
• “Compare logistic regression vs a 2-layer MLP for tabular EHR data.”

Less useful uses:

• asking for final answers without checking sources,
• pasting code blindly,
• trusting clinical claims without verification.

Best habit: ask the model to explain, then verify against the assigned source.

---

What I want you to get out of Week 1

By the end of the week, you should be able to answer:

• What is the difference between data from routine care and data from an RCT?
• What is a computable phenotype?
• Why can T2D not be identified reliably from a single field alone?
• What do age, sex/gender, race/ethnicity, ICD codes, and comorbidities contribute to a model?
• What are the main moving parts in a basic neural network pipeline: preprocessing, architecture, regularization, training, and evaluation?

---

Core packet (do these first)

1) EHRs and real-world data

Watch first

1. The Future of Health Care: Electronic Health Records (HealthIT.gov)

   • https://www.healthit.gov/video/future-health-care-electronic-health-records
   • Goal: get a fast, nontechnical picture of what EHRs are and why they matter.

2. Protecting Health Information Video (ASTP/HealthIT.gov)

   • https://www.healthit.gov/resources/video-protecting-health-information/
   • Goal: remember that health data work always sits inside privacy/security constraints.

Read next

3. FDA: Real-World Evidence

   • https://www.fda.gov/science-research/science-and-research-special-topics/real-world-evidence
   • Read only the definitions and examples.
   • Focus question: what counts as real-world data, and how is it different from evidence from a randomized trial?

4. HHS: HIPAA Privacy Rule (skim)

   • https://www.hhs.gov/hipaa/for-professionals/privacy/index.html
   • Focus question: what is protected health information (PHI), in plain language?

5. HHS: De-identification guidance (skim just the overview)

   • https://www.hhs.gov/hipaa/for-professionals/privacy/special-topics/de-identification/index.html
   • Focus question: why is “de-identified” not the same as “zero risk”?

---

2) Computable phenotyping

Read/watch

1. PheKB: What is the Phenotype KnowledgeBase?

   • https://phekb.org/
   • Read the front-page description only.
   • Goal: understand that a phenotype algorithm is a transportable definition built from data elements.

2. PheKB: Type 2 Diabetes Mellitus phenotype

   • https://www.phekb.org/phenotype/type-2-diabetes-mellitus
   • Do not try to memorize all the code lists.
   • Focus on the logic: diagnoses + meds + labs + exclusions.

3. OHDSI 2019 Cohort Definition & Phenotyping Tutorial (Part 1)

   • https://www.ohdsi.org/2019-tutorials-cohort-definition-phenotyping/
   • Watch the introductory video.
   • Goal: see how phenotype definitions are treated as explicit cohort logic rather than vague clinical intuition.

4. OHDSI Cohort Definition and Phenotyping Tutorial (2018)

   • https://www.ohdsi.org/past-events/cohort-definitionphenotyping-tutorial/
   • Optional this week; useful later for vocabulary basics and phenotype evaluation.

Write 3 bullets after this section:

• what a computable phenotype is,
• why “case” and “control” need explicit rules,
• one reason phenotype transport across sites is hard.

---

3) Type 2 diabetes basics

Read

1. CDC: Diabetes Basics

   • https://www.cdc.gov/diabetes/about/index.html
   • Goal: get the broad picture and public-health framing.

2. NIDDK: Type 2 Diabetes

   • https://www.niddk.nih.gov/health-information/diabetes/overview/what-is-diabetes/type-2-diabetes
   • Focus on: what T2D is, symptoms, causes, and why diagnosis can be delayed.

3. NIDDK: Diabetes Tests & Diagnosis

   • https://www.niddk.nih.gov/health-information/diabetes/overview/tests-diagnosis
   • Focus on the thresholds for A1C, fasting plasma glucose, OGTT, and random glucose.

4. NIDDK: The A1C Test

   • https://www.niddk.nih.gov/health-information/diagnostic-tests/a1c-test
   • Goal: understand what A1C measures and why it is useful.

5. CDC: A1C Test for Diabetes and Prediabetes

   • https://www.cdc.gov/diabetes/diabetes-testing/prediabetes-a1c-test.html
   • Quick reinforcement of thresholds.

---

4) Comorbidity, Charlson, and ICD codes

Read/watch

1. NIH glossary: Comorbidity

   • https://clinicalinfo.hiv.gov/en/glossary/comorbidity
   • 1-minute definition.

2. NCI Comorbidity Index Overview

   • https://healthcaredelivery.cancer.gov/seermedicare/considerations/comorbidity.html
   • Focus on the history of the Charlson index and why ICD-coded data are used to derive comorbidity burden.

3. CDC: ICD-10-CM

   • https://www.cdc.gov/nchs/icd/icd-10-cm/
   • Read the “What to know” section only.

4. CMS web training: Diagnosis Coding, Using the ICD-10-CM

   • https://gov-mirror.org/www.cms.gov/Outreach-and-Education/MLN/WBT/MLN6447308-ICD-10-CM/icd10cm/index.html
   • Optional this week. Do just Lesson 1 (ICD-10 Basics) if time permits.

5. AHRQ HCUP Software Tools Tutorial

   • https://hcup-us.ahrq.gov/tech_assist/software/508course.jsp
   • Optional. Useful later for seeing how ICD diagnosis codes are operationalized into comorbidity systems.

Why this matters for our project

• In your modeling dataset, many comorbidity variables are already present as patient-level binary features.
• Think of these as a compressed summary of disease burden and clinical history.

---

5) Neural networks: intuition first

Watch in this order

1. 3Blue1Brown: Neural networks topic page

   • https://www.3blue1brown.com/topics/neural-networks

2. But what is a Neural Network?

   • https://www.3blue1brown.com/lessons/neural-networks

3. Gradient descent, how neural networks learn

   • https://www.3blue1brown.com/topics/neural-networks
   • Use the topic page above to access the lesson.

4. What is backpropagation really doing?

   • Also from the same topic page.

5. Backpropagation calculus

   • Optional for this week.

What to pay attention to

• input features,
• hidden layers,
• weights and biases,
• nonlinear activation,
• loss function,
• gradient descent,
• train vs validation performance.

---

6) Neural networks in R

Pick one main path and skim the other.

Path A (recommended): torch for R

1. torch for R: Start here

   • https://torch.mlverse.org/start/

2. Guess the correlation

   • https://torch.mlverse.org/start/guess_the_correlation/
   • End-to-end example with data loading, model definition, training, and evaluation.

3. torch technical tutorial series

   • https://torch.mlverse.org/technical/
   • Great for tensors, autograd, modules, and optimizers.

Path B (also good): Keras / TensorFlow for R

1. TensorFlow for R tutorials hub

   • https://tensorflow.rstudio.com/tutorials/

2. Basic Regression

   • https://tensorflow.rstudio.com/tutorials/keras/regression
   • Good first example for tabular features.

3. Basic Image Classification

   • https://tensorflow.rstudio.com/tutorials/keras/classification.html
   • Good for learning the standard training pattern.

4. Overfit and underfit

   • https://tensorflow.rstudio.com/tutorials/keras/overfit_and_underfit.html
   • Important for dropout, L2 regularization, early stopping, and why bigger is not always better.

5. Sequential model guide

   • https://tensorflow.rstudio.com/guides/keras/sequential_model.html
   • Good reference for how to build a simple stack of dense layers.

6. Training & evaluation with built-in methods

   • https://tensorflow.rstudio.com/guides/keras/training_with_built_in_methods.html
   • Useful when you start changing callbacks and learning-rate schedules.

---

What not to over-focus on this week

Do not spend much time on:

• obscure clinical edge cases,
• memorizing code lists,
• fancy deep architectures,
• squeezing out benchmark performance.

This week is about building intuition and vocabulary.

---

Light exploratory data analysis (EDA) on the T2D dataset

When you start EDA, check these first:

1. sample size and class balance (i.e., prevalence of T2D),
2. age distribution,
3. counts by gender / race / ethnicity,
4. prevalence of each comorbidity feature,
5. missingness or odd category values,
6. duplicate rows or identifier issues,
7. whether any text-like medication column should be dropped for the first model.

For this rotation, demographics are not the hard concept. They are mostly an EDA / encoding issue. The harder concept is how demographics, comorbidities, and phenotype rules interact.

---

Optional R packages to know about later

These are not required for Week 1, but useful later:

• arrow for parquet files
• tidyverse for data wrangling and visualization
• recipes / rsample / yardstick for preprocessing and evaluation
• torch and luz for neural networks in R
• keras3 for simple deep learning in R
• comorbidity for computing Charlson / Elixhauser scores from ICD data