Applications & Ethics

ML in the Real World

🔄 Lesson 6 covered the tools — scikit-learn, PyTorch, TensorFlow, and the Python stack. Now let’s see what people actually build with them, and the ethical challenges that come with deploying ML at scale.

Healthcare: Where ML Saves Lives

Healthcare ML is projected to grow from $37 billion to $614 billion by 2034. The applications are already changing how medicine works.

Medical imaging: ML models analyze X-rays, MRIs, and CT scans — sometimes catching things radiologists miss. Google’s DeepMind built a model that detects over 50 eye diseases from retinal scans as accurately as world-leading ophthalmologists. The model doesn’t replace doctors. It flags cases that need urgent attention, reducing the time between scan and treatment.

Drug discovery: Traditional drug development takes 10-15 years and costs $2.6 billion on average. ML models predict which molecular compounds are likely to be effective, cutting early-stage screening from years to weeks. Insilico Medicine used ML to identify a drug candidate for idiopathic pulmonary fibrosis in 18 months — a process that typically takes 4-5 years.

Predictive diagnostics: ML models analyze patient records to predict who’s likely to develop conditions like diabetes, heart disease, or sepsis — before symptoms appear. Early intervention saves lives and reduces healthcare costs.

✅ Quick Check: A hospital wants to predict which ER patients will need ICU admission within 24 hours, using vital signs and lab results from their structured medical records. Based on what you learned in Lesson 3, which algorithm family would you start with? Tree-based models (random forest or XGBoost) — this is structured tabular data with a classification task. Start simple, then add complexity only if needed. A neural network would be overkill for this type of data.

Finance: Where ML Catches Fraud

Fraud detection: Credit card companies process millions of transactions daily. ML models flag suspicious patterns in real-time — unusual locations, spending patterns, or timing that deviate from a customer’s normal behavior. Mastercard’s ML system evaluates 143 billion transactions annually, reducing false declines by 50%.

Algorithmic trading: ML models analyze market data, news sentiment, and economic indicators to make trading decisions in milliseconds. Hedge funds like Renaissance Technologies and Two Sigma built their entire business models around ML-driven trading.

Credit scoring: Traditional credit scores use a handful of variables. ML models consider hundreds of features — payment patterns, spending behavior, employment stability — to more accurately predict default risk. This can expand credit access to “thin file” applicants who lack traditional credit history.

Marketing: Where ML Predicts Behavior

Recommendation engines: Netflix estimates its recommendation system saves $1 billion annually by reducing churn. When people find content they enjoy, they stay subscribed. Amazon attributes 35% of purchases to its recommendation engine.

Customer segmentation: Remember K-means clustering from Lesson 3? Companies use it to discover natural customer groups from purchase data — then tailor marketing, pricing, and products to each segment.

Churn prediction: Telecom companies, SaaS businesses, and subscription services use ML to predict which customers are about to leave. A model trained on historical churn data can flag at-risk customers weeks before they cancel, giving retention teams time to intervene.

✅ Quick Check: Spotify uses ML to create personalized playlists. What type of ML is this? Supervised learning — the model learns from your listening history (labeled data: songs you played, skipped, saved) to predict which new songs you’ll enjoy. It’s a recommendation system using collaborative filtering (finding users with similar taste) and content-based filtering (analyzing audio features of songs you like).

The Ethics Problem

ML is powerful. It’s also dangerous when deployed without careful thought about fairness, bias, and accountability.

The core issue: ML models learn from historical data. If that data reflects historical discrimination, the model replicates it — and often amplifies it. The model isn’t “biased” in a human sense. It found a pattern in the data and optimized for it. But the real-world impact is the same.

Real cases that went wrong:

• Amazon’s hiring tool (2018): Trained on 10 years of resumes, the model learned to penalize resumes containing the word “women’s” (as in “women’s chess club”) because the historical hiring data skewed male. Amazon scrapped it.
• COMPAS recidivism scores: Used by US courts to predict reoffending risk. A ProPublica investigation found the system was twice as likely to falsely label Black defendants as high-risk compared to white defendants.
• Healthcare algorithm (2019): Used by major US hospitals, it systematically deprioritized Black patients for follow-up care because it used healthcare spending (not health needs) as the training signal.

Fairness: Not Just “Don’t Be Biased”

Fairness in ML is harder than it sounds because there are multiple definitions of fairness — and they’re mathematically incompatible.

You can’t satisfy all three simultaneously (proven mathematically). The choice depends on context — which type of unfairness is least acceptable for this specific application?

Explainable AI (XAI)

When an ML model denies someone a loan, rejects a job application, or flags a medical scan, people deserve to know why. This is where Explainable AI (XAI) comes in.

The problem: Neural networks and complex ensemble models are “black boxes.” They produce predictions but can’t explain their reasoning in human terms.

The solutions:

• LIME (Local Interpretable Model-agnostic Explanations): Explains individual predictions by approximating the model locally with a simpler, interpretable model
• SHAP (SHapley Additive exPlanations): Shows how much each feature contributed to a specific prediction
• Inherently interpretable models: Decision trees and linear regression are transparent by design — you can trace every prediction

The regulatory push: The EU AI Act (enforced 2025) requires transparency and human oversight for high-risk AI systems — including credit scoring, hiring, healthcare, and law enforcement. Organizations deploying ML in these areas must be able to explain how decisions are made.

✅ Quick Check: Your model uses 200 features to predict employee attrition. HR asks: “Why did the model flag this employee as likely to leave?” Which XAI tool would help? SHAP — it shows the contribution of each feature to this specific prediction. You might find that this employee’s combination of low salary growth, high overtime hours, and recent manager change drove the prediction. SHAP makes the “why” visible.

Key Takeaways

• Healthcare ML is transforming diagnostics, drug discovery, and patient care — market growing from $37B to $614B by 2034
• Finance ML powers fraud detection, trading, and credit scoring — processing billions of transactions in real-time
• Marketing ML drives personalization, recommendations, and churn prediction — Netflix saves $1B/year from recommendations alone
• ML models replicate and amplify biases in training data — Amazon’s hiring tool, COMPAS, healthcare algorithms all showed discriminatory outcomes
• Fairness has multiple definitions (demographic parity, equal opportunity, predictive parity) that are mathematically incompatible — choose based on context
• XAI tools (LIME, SHAP) make black-box predictions interpretable — increasingly required by regulation (EU AI Act)

Up Next

You’ve covered concepts, algorithms, data, evaluation, tools, applications, and ethics. Lesson 8 brings it all together — a capstone that helps you design your ML learning path, choose your first project, and build a practical foundation for continued growth.