Scikit-Learn Algorithm Cheat Sheet — Clinical Epidemiology Edition: Choosing the Right ML Type

🟡 START

You begin by asking:

What kind of problem am I trying to solve in my study?

Common scenarios in clinical epidemiology:

• Identifying patients with a disease → diagnosis

• Predicting future outcomes → prognosis

• Estimating biomarkers or lab values → continuous prediction

• Finding hidden patient subgroups → phenotyping

• Simplifying many variables into a few patterns → data reduction
  

---

🔹 1. Do you have labeled data?

• YES → Go to Supervised Learning (you know the outcome)

  • e.g., “We know who died / survived, who has diabetes / doesn’t.”

• NO → Go to Unsupervised Learning

  • e.g., “We don’t have labels, just want to find patient subgroups.”
    

---

🟩 2. Supervised Learning

If you have labeled data, your goal is to predict an outcome.

What are you predicting?

• A category (classification) → e.g., disease status, ICU admission (Yes/No)

• A quantity (regression) → e.g., blood pressure, hospital stay length, cost
  

---

🟥 3. Classification (Predicting a Category)

Goal: Predict discrete outcomes — e.g., disease / no disease, high-risk / low-risk.

a. Do you have text data (e.g., clinical notes, discharge summaries)?

• YES → Use Naive Bayes 📘 Example: Automatically detect mentions of pneumonia in hospital discharge notes.

• NO → Continue.
  

b. <100K samples?

• YES → Try Linear SVC 📘 Example: Predict 30-day readmission (Yes/No) in a cohort of 10,000 patients.

• NO → Try SGD Classifier 📘 Example: Predict sepsis onset from millions of EHR time-points in ICU data.
  

c. Are you using kernel approximation (non-linear patterns)?

• YES → Try SVC (Support Vector Classifier) 📘 Example: Detect diabetic retinopathy (yes/no) from imaging data with non-linear relationships.

• NO → Try KNeighbors Classifier or Ensemble Classifiers (Random Forest, Gradient Boosting) 📘 Example: Predict who will develop heart failure after myocardial infarction using demographics + labs.

---

🟦 4. Regression (Predicting a Quantity)

Goal: Predict a continuous variable — e.g., lab value, survival time, hospital stay.

a. More than 50 samples?

• NO → Get more data. 📘 Example: You can’t predict HbA1c change from 10 patients; you need more cases.

• YES → Continue.
  

b. <100K samples?

• YES → Continue. Most clinical studies fit here.

• NO → Try SGD Regressor 📘 Example: Predict total hospital charges using millions of EHR records.
  

c. Should few features be important?

• YES → Try Lasso or ElasticNet 📘 Example: Identify key lab markers predicting survival time after cancer treatment.(Lasso can shrink unimportant predictors to zero → easier interpretation.)

• NO → Try Ridge Regression or SVR(kernel='linear') 📘 Example: Predict systolic BP from many moderately correlated features.

• If non-linear patterns present: Try SVR(kernel='rbf') or Ensemble Regressors 📘 Example: Predict ICU length-of-stay from complex physiological data.
  

---

🟪 5. Unsupervised Learning

If you don’t have labeled outcomes, you’re exploring hidden structures or relationships.

Are you just exploring patterns?

• YES → Dimensionality Reduction 📘 Example: Visualize patient clusters based on 200 lab markers in 2D.

• NO → Clustering 📘 Example: Identify unknown disease phenotypes among patients with chronic kidney disease.
  

---

🟨 6. Dimensionality Reduction

Goal: Simplify large feature sets into fewer meaningful components.

a. <10K samples?

• YES → Try Randomized PCA 📘 Example: Summarize 100 lab variables into 3 principal components explaining most variance.

• NO → Continue.
  

b. Using kernel approximation (non-linear data)?

• YES → Try Isomap or Spectral Embedding 📘 Example: Explore nonlinear patterns among genetic expression data.

• NO → Try LLE (Locally Linear Embedding) 📘 Example: Reduce multi-omics data to visualize patient progression patterns.
  

---

🟫 7. Clustering

Goal: Group similar patients (unsupervised).

a. <10K samples?

• YES → Continue. 📘 Example: 5,000 heart failure patients — group by comorbidity profiles.

• NO → Try MiniBatch KMeans 📘 Example: Cluster 1 million insurance records by utilization patterns.
  

b. Number of categories known?

• YES → Try K-Means 📘 Example: Group patients into 3 severity levels (mild/moderate/severe).

• NO → Continue.
  

c. <10K samples?

• YES → Try MeanShift or VBGMM (Variational Gaussian Mixture) 📘 Example: Identify latent phenotypes in sepsis using vitals and labs (without predefining the number of clusters).

• NO → Tough luck 😅 (Try sampling smaller subsets or feature reduction first.)
  

---

🔁 Summary by Problem Type (Clinical Examples)

Problem Type	Algorithm	Clinical Epidemiology Example
Classification	Naive Bayes, Linear SVC, KNeighbors, SGD, SVC, Ensemble	Predict diabetes diagnosis, detect pneumonia in radiology notes
Regression	Lasso, ElasticNet, Ridge, SVR, Ensemble	Predict HbA1c, BP, or survival time
Clustering	KMeans, MeanShift, VBGMM	Find phenotypes of chronic kidney disease or sepsis subtypes
Dim. Reduction	PCA, Isomap, LLE, Spectral Embedding	Summarize 100 biomarkers into 2 latent factors
Text Data (special)	Naive Bayes, SVM	Detect adverse drug events in clinical notes

---

Practical Clinical Examples — Choosing the Right ML Type

Clinical Question	ML Type	Example Algorithm(s)	Why / Notes
🩺 “Can we predict which hypertensive patients will develop stroke?”	Classification	RandomForestClassifier, LogisticRegression, SupportVectorClassifier (SVC)	Outcome is categorical (Yes/No). Models predict probability of stroke occurrence.
🏥 “Which factors best predict hospital stay length?”	Regression	RandomForestRegressor, Lasso, ElasticNet, RidgeRegression	Outcome is continuous (days). Regularized regressions highlight key predictors; Random Forest can handle non-linear relationships.
💨 “Can we find subgroups of COPD patients with similar symptoms?”	Clustering (Unsupervised)	KMeans, GaussianMixtureModel (GMM), MeanShift, VBGMM	No labels — the goal is to discover hidden phenotypes or patient subgroups.
🧬 “Can we reduce 300 lab markers into a few composite risk factors?”	Dimensionality Reduction	PCA, Isomap, t-SNE, LLE, SpectralEmbedding	Reduce many correlated biomarkers to fewer components for visualization or downstream analysis.
📝 “Can we automatically identify mentions of adverse drug events in notes?”	Text Classification (NLP)	NaiveBayes, LinearSVC, LogisticRegression (with TF-IDF)	Text data with labeled outcomes; algorithms classify notes mentioning adverse drug events.
🧠 “Can we predict which ICU patients will deteriorate within 24 hours?”	Classification (Time-series)	GradientBoostingClassifier, XGBoost, LSTM (if temporal)	Binary outcome (deteriorate vs stable). Uses time-based features from EHR.
💵 “Can we estimate total hospital cost per patient?”	Regression	RandomForestRegressor, XGBoostRegressor	Continuous numeric outcome (cost). Captures nonlinear effects of demographics and clinical variables.
⚕️ “Can we group diabetic patients by treatment response patterns?”	Clustering (Phenotyping)	KMeans, Hierarchical Clustering, DBSCAN	No known labels. Unsupervised grouping to discover response phenotypes.
🧩 “Can we visualize overall patient similarity based on 100 variables?”	Dimensionality Reduction (Visualization)	PCA, UMAP, t-SNE	Helps visualize complex data in 2D/3D space — e.g., to show clusters of similar patients.
💊 “Does a specific drug reduce mortality in COVID-19 patients?”	Causal Inference (not pure ML)	PropensityScoreMatching, TargetedMaximumLikelihood (TMLE), Double Machine Learning	Seeks causal effect, not prediction — integrates ML for confounder adjustment.

⚕️ How This Helps in Clinical Epidemiology

• ✅ Match your question → ML family → right algorithm

• 🔍 Keeps interpretability and clinical relevance central

• 📈 Encourages external validation (across hospitals or populations)

• 📊 Use feature importance or SHAP values for transparency

• ⚠️ Remember: Prediction ≠ Causation — causal inference requires different frameworks