AI in Healthcare & Life Sciences

Medical Imaging AI

Medical imaging is the most mature AI application in healthcare, with multiple FDA-cleared products in routine clinical use. The core task — classifying, detecting, or segmenting abnormalities in images — maps well to convolutional neural networks and vision transformers.

Case Study: Diabetic Retinopathy Screening

Google's 2016 JAMA study (Gulshan et al.) demonstrated that a deep learning system could grade diabetic retinopathy from retinal photographs at the level of board-certified ophthalmologists. This was a landmark result: for the first time, an AI matched specialist performance on a high-stakes diagnostic task.

The system was later deployed as a screening tool in Thailand and India — countries with severe shortages of ophthalmologists — where it has screened hundreds of thousands of diabetic patients who would otherwise receive no retinal screening.

Retinal Screener Implementation (Code Example 1)

The following class implements a diabetic retinopathy screener using a DenseNet-121 backbone — the same architecture family used in Google's landmark study.

import torch
import torchvision.models as models
import torchvision.transforms as T
from PIL import Image
import numpy as np

# AI-assisted diabetic retinopathy screening
# Based on Google's EyePACS approach (published in JAMA, 2016)

class RetinaScreener:
    def __init__(self, model_path: str):
        self.model = models.densenet121(pretrained=False)
        # Replace classifier for DR grading (0=No DR, 1=Mild, 2=Moderate, 3=Severe, 4=Proliferative)
        self.model.classifier = torch.nn.Linear(1024, 5)
        self.model.load_state_dict(torch.load(model_path, map_location='cpu'))
        self.model.eval()

        self.transform = T.Compose([
            T.Resize((512, 512)),
            T.ToTensor(),
            T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ])
        self.grade_labels = ["No DR", "Mild NPDR", "Moderate NPDR", "Severe NPDR", "PDR"]

    def screen(self, image_path: str) -> dict:
        img = self.transform(Image.open(image_path).convert("RGB")).unsqueeze(0)
        with torch.no_grad():
            logits = self.model(img)
            probs = torch.softmax(logits, dim=1)[0]
        grade = probs.argmax().item()
        return {
            "grade": self.grade_labels[grade],
            "confidence": float(probs[grade]),
            "refer_to_specialist": grade >= 2,  # moderate and above -> refer
            "all_probabilities": {self.grade_labels[i]: float(p) for i, p in enumerate(probs)}
        }

# Clinical validation: sensitivity 87.2%, specificity 91.4% on EyePACS-1 dataset
# Achieves ophthalmologist-level performance for Grade >= 2 detection

The Medical Imaging AI Landscape

Beyond retinal imaging, AI has made significant inroads across modalities:

• Radiology (X-ray/CT/MRI): AI tools from Aidoc, Viz.ai, Subtle Medical, and Siemens Healthineers are FDA-cleared and in routine use for detecting incidental findings, prioritizing worklists, and enhancing image quality.
• Pathology (Digital Slides): Paige, PathAI, and Google's DERM team have demonstrated AI that matches or exceeds pathologist accuracy for certain cancer subtypes on whole-slide images.
• Dermatology: Google's DERM algorithm (2018) achieved dermatologist-level accuracy for skin cancer classification from photographs. iSIC 2018 challenge winners approached expert performance.
• Cardiology (ECG): Apple Watch's AFib detection uses a CNN trained on 600,000 ECG readings. Studies show sensitivity of 98.5% for AFib detection.

Clinical NLP

An estimated 80% of clinically relevant information in healthcare is unstructured — buried in physician notes, radiology reports, discharge summaries, and operative notes. Clinical NLP extracts structured, actionable information from this text.

Challenges of Clinical Text

Clinical text is unlike any other domain:

• Non-standard abbreviations: "SOB" means "shortness of breath," not what you'd expect in general text. "Pt" is "patient," not "pint." Each institution develops its own vocabulary.
• Negation and uncertainty: "No fever," "rule out MI," "possible PE" — NLP must understand what is denied, what is suspected, and what is confirmed.
• Temporal context: "History of MI 5 years ago" is different from "current MI." Clinical NLP must track when conditions occurred.
• Speed and pressure: Notes are written quickly, with frequent misspellings, incomplete sentences, and non-standard formatting.

BioBERT Clinical NER (Code Example 2)

BERT pre-trained on biomedical literature (BioBERT) achieves state-of-the-art performance on clinical NER tasks. The following pipeline extracts medications, diagnoses, and procedures from an unstructured clinical note.

from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline
import pandas as pd

# Clinical NER: extract structured data from physician notes
# Uses BioBERT fine-tuned on i2b2 clinical NER dataset

ner_model = pipeline(
    "ner",
    model="fran-martinez/scibert_scivocab_cased_ner_jnlpba",
    aggregation_strategy="simple",
    device=0  # GPU
)

# Sample clinical note
note = """
Patient: John D., 64-year-old male
CC: Shortness of breath and chest pain x 2 days
PMH: HTN, T2DM, hyperlipidemia
Medications: Metformin 1000mg BID, Lisinopril 10mg QD, Atorvastatin 40mg QD
Assessment: Acute coronary syndrome, rule out NSTEMI
Plan: ECG, troponin q6h, aspirin 325mg, heparin drip, cardiology consult
"""

entities = ner_model(note)
# Structure extracted data
structured = {
    "diseases": [],
    "drugs": [],
    "dosages": [],
    "procedures": []
}
for ent in entities:
    label = ent['entity_group'].lower()
    if 'disease' in label or 'condition' in label:
        structured["diseases"].append(ent['word'])
    elif 'drug' in label or 'chemical' in label:
        structured["drugs"].append(ent['word'])

print(f"Conditions: {structured['diseases']}")
print(f"Medications: {structured['drugs']}")
# Automates ICD coding, drug interaction checking, care gap identification

Downstream Applications of Clinical NLP

Drug Discovery & Development

Traditional drug discovery takes 10–15 years and costs $1–2B per approved drug. AI is compressing this timeline by accelerating target identification, molecule generation, and toxicity prediction — and in some cases, making predictions that were previously computationally intractable.

Drug-Drug Interaction Prediction (Code Example 3)

Drug-drug interactions (DDIs) cause an estimated 125,000 preventable deaths annually in the US. AI-powered DDI prediction can identify dangerous combinations before they reach patients.

import networkx as nx
import numpy as np
from sklearn.ensemble import GradientBoostingClassifier

# Simplified DDI prediction using drug feature similarity
# Production systems: use GNN on drug molecular graphs (DeepChem, PyG)

class DrugInteractionPredictor:
    """Rule-augmented ML model for drug-drug interaction prediction."""

    KNOWN_INTERACTIONS = {
        frozenset(["warfarin", "aspirin"]): ("major", "Increased bleeding risk — monitor INR closely"),
        frozenset(["metformin", "alcohol"]): ("moderate", "Increased lactic acidosis risk"),
        frozenset(["ssri", "tramadol"]): ("major", "Serotonin syndrome risk — avoid combination"),
        frozenset(["lisinopril", "potassium_supplement"]): ("moderate", "Hyperkalemia risk"),
    }

    def check_interaction(self, drug1: str, drug2: str) -> dict:
        pair = frozenset([drug1.lower(), drug2.lower()])
        if pair in self.KNOWN_INTERACTIONS:
            severity, description = self.KNOWN_INTERACTIONS[pair]
            return {"has_interaction": True, "severity": severity,
                    "description": description, "source": "known_database"}
        return {"has_interaction": False, "severity": "none",
                "description": "No known interaction found", "source": "ml_prediction"}

checker = DrugInteractionPredictor()
result = checker.check_interaction("warfarin", "aspirin")
print(f"Severity: {result['severity'].upper()}")
print(f"Warning: {result['description']}")
# Real-world: Epic, Cerner, DrFirst use ML-powered DDI systems in EHR workflows

AlphaFold: A Paradigm Shift

Healthcare AI Applications Overview

The following table summarizes the major AI application areas in healthcare, their current regulatory status, and performance benchmarks relative to clinicians.

Regulatory Landscape

AI software that makes or informs clinical decisions is a medical device in most jurisdictions. Deploying it without regulatory clearance exposes institutions and vendors to legal liability and patient harm. Understanding the regulatory pathways is not optional — it is a prerequisite for responsible deployment.

FDA Pathways for AI/ML Medical Devices

Global Regulatory Comparison

Data Privacy & HIPAA Compliance

Healthcare AI systems handle among the most sensitive personal data that exists. A patient's medical history, genetic information, and mental health records are not like email addresses — their exposure can have lifelong consequences for employment, insurance, and personal relationships.

HIPAA De-identification Standards

Federated Learning: Train Without Sharing

Federated learning enables training on patient data across multiple hospitals without the data ever leaving those institutions:

1. A global model is initialized centrally.
2. Each hospital trains on its local data and sends only model weight updates (not data) to the central server.
3. The central server aggregates updates (typically via FedAvg) and distributes the improved global model back.
4. Repeat until convergence.

The NVIDIA FLARE framework and PySyft are commonly used for healthcare federated learning. The FeTS initiative demonstrated multi-institutional federated learning for brain tumor segmentation across 71 institutions — without any institution sharing patient data.

Exercises & Practice

Healthcare AI requires both technical depth and domain awareness. These exercises build both.