Lesson 1: AI in Healthcare — Overview & Ethics

AI in Health & Healthcare: Real Battle Applications

Part 1: Medical AI Platform & Medical Data

xdev.asia

This article lays the thinking foundation for the entire series. You'll understand why AI works differently in healthcare than in other industries — and why it's more important than any algorithm.

---

1. Why is AI in Healthcare completely different?

When building an AI to classify cat/dog photos, 5% error is okay. When building AI to diagnose lung cancer, a 5% error means thousands of people being missed or receiving unnecessary chemotherapy.

Three things make healthcare AI different from every other domain:

1.1. Stakes are extremely high

Domain	Consequences if AI is wrong
Movie Recommendation	User is not good at watching movies
Spam filter	Email goes to wrong folder
Medical diagnosis	Patient death or disability
Drug dosage AI	Overdose or underdose

This is not a reason to fear AI in healthcare — but a reason to do more right, more carefully.

1.2. Data is both precious and scarce

Precious: Each patient is a natural "experiment" that cannot be replicated. Rare cancer data may only have a few hundred cases globally.

Khan: HIPAA (US), GDPR (EU), Decree 13/2023 (Vietnam) — health data is most protected across all industries. Cannot be scraped from the internet.

Imbalanced: 99% of patients are healthy, 1% have rare diseases → naïve model only predicts "healthy" which is correct 99% of the time but is useless.

1.3. "Ground truth" does not completely exist

In NLP, correct translations are verifiable. In medical imaging:

• Inter-rater reliability: 3 radiologists reading the same film can give 3 different results
• Gold standard is usually the result of surgery or biopsy — expensive, slow, invasive
• The disease can progress: the label is correct today, wrong 6 months later

---

2. Landscape: Where is AI being used in Healthcare?

AI Healthcare Ecosystem
│
├── 🔬 Chẩn đoán & Phát hiện bệnh
│   ├── Radiology AI (X-ray, CT, MRI)         ← Mature, FDA-cleared products
│   ├── Pathology AI (digital slides)          ← Growing fast
│   ├── Dermatology AI (skin lesions)          ← Consumer + clinical
│   ├── Ophthalmology (retinal scan)           ← DeepMind, Google leading
│   └── Clinical Decision Support Systems     ← Hospital workflow tools
│
├── 💊 Drug Discovery & Development
│   ├── Target identification                  ← Bioinformatics + ML
│   ├── Molecule design (generative AI)        ← GNN, Transformer
│   ├── ADMET prediction                       ← Replace wet lab screening
│   └── Clinical trial optimization            ← Patient matching, protocol
│
├── 🧬 Genomics & Precision Medicine
│   ├── Variant calling                        ← DeepVariant (Google)
│   ├── Polygenic risk scores                  ← GWAS + ML
│   └── Pharmacogenomics                       ← Drug-gene interactions
│
├── 📋 EHR & Clinical Operations
│   ├── Clinical NLP (notes → structured)     ← Extract diagnoses, meds
│   ├── ICD coding automation                  ← Billing, compliance
│   ├── Readmission prediction                 ← Care management
│   └── Sepsis early warning                   ← ICU monitoring
│
├── ⌚ Remote Monitoring & Wearables
│   ├── AFib detection (ECG/PPG)               ← Apple Watch cleared
│   ├── Continuous glucose monitoring + AI    ← Diabetes management
│   └── Mental health monitoring              ← Sleep, activity patterns
│
└── 🏥 Operations & Administration
    ├── Medical coding + billing               ← Revenue cycle management
    ├── Appointment no-show prediction         ← Scheduling optimization
    └── Supply chain (drug demand forecast)   ← Hospital logistics

2.1. Maturity level by application

Application	Phase	Representative
Retinal screening (DR)	Production	IDx-DR (FDA cleared 2018)
Chest X-ray AI	Production	Aidoc, Viz.ai
Sepsis prediction	Clinical deployment	Epic Sepsis Model
Drug molecular design	Research → Early clinical	Insilico Medicine
Surgical robot AI	Experimental	Intuitive Surgical R&D
Digital twin patient	Research	2030+ horizon

---

3. Lifecycle of a Medical AI Project

Understanding this lifecycle helps you ask the right questions from the beginning:

1. Problem Definition
   "Bài toán lâm sàng là gì? AI có phải là giải pháp phù hợp không?"
   ↓
2. Data Strategy
   "Dữ liệu nào có? Privacy? Annotation cost? Class imbalance?"
   ↓
3. Model Development
   "Architecture nào? Baseline metrics? Validation methodology?"
   ↓
4. Clinical Validation
   "Prospective study hay retrospective? Comparison với clinicians?"
   ↓
5. Regulatory Submission
   "FDA 510(k)? CE Mark? SaMD class?"
   ↓
6. Clinical Integration
   "Workflow integration? EHR compatibility? Change management?"
   ↓
7. Post-market Monitoring
   "Performance drift? Adverse events? Retraining strategy?"

Bloody lesson: 85% of medical AI projects fail not because of bad algorithms — but because steps 1, 6, 7 are skipped.

---

4. AI Ethics in Healthcare

4.1. The four traditional pillars of medical ethics → apply to AI

Principles of medical ethics	Meaning	Apply to AI
Beneficence (benefit)	Acting for the benefit of patients	AI must improve outcomes, not just optimize metrics
Non-maleficence (not harmful)	"First, do no harm"	Understand AI failure modes before deploying
Autonomy (autonomy)	Patients have the right to decide	Informed consent when AI participates in diagnosis
Justice (fairness)	Treat all patients fairly	Fairness across demographic groups

4.2. Three unique ethical issues of Medical AI

Issue 1: Bias and Health Disparities

AI does not invent bias — it amplifies bias that already exists in the data.

Real-life example: A 2019 study in Science (Obermeyer et al.) found that a health management algorithm used by >200 million Americans has severe bias:

Thuật toán dùng "chi phí y tế trong quá khứ" để dự đoán "mức độ bệnh"
→ Người da đen có chi phí y tế thấp hơn (vì thiếu tiếp cận dịch vụ)
→ Thuật toán kết luận họ ít bệnh hơn thực tế
→ Bệnh nhân da đen phải bệnh nặng hơn 40% mới được chăm sóc tương đương

Issue 2: Accountability — Who is responsible?

Bác sĩ tin AI → AI sai → Bệnh nhân tổn hại
   ↑
Ai có lỗi?
  - Bác sĩ không kiểm tra?
  - Nhà sản xuất AI?
  - Bệnh viện mua AI không validate?
  - FDA đã approve nhưng AI vẫn sai?

Currently: There is no clear legal framework. The EU AI Act is trying to deal with "high-risk AI" classification.

Issue 3: Automation Bias

Doctors tend to trust AI more when it provides a high confidence level:

# Nghiên cứu: Radiologists với AI assistance
# Kịch bản 1: AI đúng, bác sĩ gốc sai → bác sĩ theo AI (tốt)
# Kịch bản 2: AI sai, bác sĩ gốc đúng → bác sĩ CŨNG theo AI (nguy hiểm)
# → "Automation bias" làm bác sĩ kém hơn khi có AI kém!

Solution: The UI design does not display the confidence score before the physician reads it independently.

4.3. Framework for evaluating AI Ethics in Healthcare

# Model Card — tài liệu tối thiểu khi publish medical AI
model_card = {
    "model_details": {
        "name": "ChestAI v2.1",
        "type": "Multi-label chest X-ray classifier",
        "conditions_detected": ["Pneumonia", "Effusion", "Cardiomegaly", ...],
        "intended_use": "Screening support for radiologists, NOT standalone diagnosis"
    },
    "training_data": {
        "dataset": "CheXpert + NIH ChestX-ray14",
        "size": "224,316 images",
        "demographics": {
            "age": "18-90 years",
            "sex": "51% male, 49% female",
            "race": "WARNING: Race not captured in original dataset"
        }
    },
    "performance": {
        "overall_auc": 0.889,
        "by_condition": {"Pneumonia": 0.768, "Effusion": 0.925, ...}
    },
    "fairness_evaluation": {
        "sex_auc_gap": 0.023,  # Female vs Male
        "age_auc_gap": 0.041,  # Elderly (>70) vs Adult (30-70)
        "WARNING": "Not evaluated on non-US populations"
    },
    "limitations": [
        "NOT validated for pediatric (<18) patients",
        "Performance may degrade on non-standard beam projections",
        "Not tested on COVID-19 pneumonia"
    ],
    "out_of_scope": [
        "Final diagnosis without physician review",
        "Emergency triage decision making"
    ]
}

---

5. Regulatory Landscape: Global game rules

5.1. FDA — Software as Medical Device (SaMD)

FDA classifies AI/ML software according to risk level:

Class	Risk	Example	Road
Class I	Low	Administrative tools, low-risk support	General Controls
Class II	Average	X-ray AI (screening support)	510(k) clearance
Class III	Cao	AI decides automatic treatment	PMA (Premarket Approval)

510(k): Demonstrated "substantially equivalent" with cleared predicate device.
PMA: Must demonstrate safety and effectiveness — similar to the drug approval process.

Pre-Determined Change Control Plan (PCCP) — new regulations 2023: Allow AI model to self-update/improve within approved scope without resubmitting. It's important because medical AI needs to constantly retrain with new data.

5.2. EU AI Act + MDR

The EU AI Act (effective 2024) classifies medical AI as High-Risk AI:

• Transparency obligation: must notify the user who is interacting with the AI
• Human oversight: there must be an override mechanism
• Accuracy/robustness/cybersecurity requirements
• Conformity assessment required before entering the EU market

5.3. Vietnam

• Decree 98/2021 on medical equipment management
• AI diagnostic software: classify medical devices A/B/C/D
• Completing the legal framework for medical AI (2024-2026)
• Reality: many AI products are operating as "physician support" to avoid device classification

---

6. AI is not Silver Bullet — When NOT to use AI?

This is the most important question a medical AI engineer must answer honestly:

Do not use AI when:

1. Simple rule-based system is enough: If the problem can be solved with 5 clear if-else with 100% accuracy, do not train the model.

2. Dataset is too small and cannot be expanded: Super rare disease with 50 cases → AI will overfit, not generalize. Rare disease approaches (few-shot, synthetic data) should be considered.

3. Lack of reliable ground truth: If inter-rater agreement < 70%, the model will learn noise.

4. Can't explain to the doctor: Some clinical contexts (court, insurance) need AI that can explain each decision.

5. Regulatory pathway is unclear: Deploying the tool without knowing the SaMD class may lead to withdrawal.

---

7. Summary & Preparation for Lesson 2

After this article, you should understand:

• ✅ Medical AI is different from regular AI in stakes, data constraints, ground truth uncertainty
• ✅ Landscape applications from mature (retinal AI) to experimental (digital twin)
• ✅ The medical AI project life cycle includes 7 steps — failures are usually in steps 1, 6, 7
• ✅ Ethics: bias, accountability, automation bias
• ✅ Regulatory: FDA SaMD classes, 510(k)/PMA, EU AI Act, VN context

Lesson 2 will delve into the most important thing about medical AI: data. You'll learn how DICOM images are stored, what HL7 FHIR is, and why de-identification isn't as simple as you think.

---

Exercise

1. Find an AI medical device that has been cleared by the FDA 510(k) database. Read the predicate comparison — which device does it compare to?

2. Read the paper "Dissecting racial bias" (Obermeyer et al., 2019). Why is using "cost" as a proxy for "health" problematic?

3. For the following problem, decide: AI or rule-based? Explain why.

• Warning when blood pressure > 180/120
  • Predict whether the patient will be readmitted to the hospital within 30 days
  • Classify ECG as normal or abnormal

---

2. Architecture & Principles

Core Architecture

# Example implementation
import torch
import torch.nn as nn

class ExampleModel(nn.Module):
    def __init__(self, input_dim, output_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(input_dim, 256),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(256, 128),
            nn.ReLU(),
            nn.Linear(128, output_dim),
        )
    
    def forward(self, x):
        return self.net(x)

---

3. Practice

Setup

pip install torch transformers datasets

Training Pipeline

# Training loop
model = ExampleModel(input_dim=768, output_dim=10)
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
criterion = nn.CrossEntropyLoss()

for epoch in range(10):
    for batch in train_loader:
        optimizer.zero_grad()
        outputs = model(batch["input"])
        loss = criterion(outputs, batch["label"])
        loss.backward()
        optimizer.step()

---

4. Best Practices

Aspect	Recommendation
Data	Quality over quantity
Model	Start simple, scale up
Training	Monitor loss curves
Evaluation	Use appropriate metrics

---

Summary

Concepts	Key Takeaway
Architecture	Suitable for the problem
Training	Careful hyperparameter tuning
Evaluation	Multiple metrics