Dr. Newton & MIT Credentials Dr. Newton holds MIT certification from the Abdul Latif Jameel Clinic for Machine Learning in Health, pioneering AI healthcare applications including Mirai breast cancer detection platforms.

AI in Healthcare (Breast Cancer Focus) AI revolutionizes healthcare through enhanced disease detection and diagnostic accuracy. Studies show AI-supported mammography increases cancer detection by 17.6%, reduces interval cancers by 30%, and outperforms traditional systems.

Florida Breast Cancer Early Detection Programs Key programs include Florida's Mary Brogan Program (free/low-cost screening), Sylvester Cancer Center, Moffitt Cancer Center, and county health departments offering comprehensive screening services statewide.

This is AI Generated...

AI can facilitated education and teaching at CONFERENCES, COURSES and TRAINING for CAREGIVERS, DOCTORS, NURSES, & HEALTH ADVOCATES.

THIS IMAGE IS ACCURATE ... BUT SOMETIMES AI MAKES SIGNIFICANT ERRORS.

CORRECT IMAGE

INCORRECT IMAGE

THE ABOVE IMAGES ARE NOT ACCURATE: 2 main issues:

Spelling/Labeling Error = It says “Verbera Body” instead of “Vertebral Body.” This could confuse learners since accuracy in anatomy labeling is critical.
Spinal Cord Labeling = The spinal cord typically ends around L1–L2 in adults. Below that, the canal contains the cauda equina (nerve roots) rather than cord. In the lumbar region where herniations usually occur, this image should reference nerve roots rather than the spinal cord itself.

👉 For teaching caregivers and advocates, correcting those terms would make the illustration clearer and medically accurate.

Conversational | Intake AI | Ex. Work Flow

1. Emitrr – AI Chatbot & Virtual Receptionist

Use: Patient intake, reminders, scheduling, two-way messaging.
HIPAA Features: End-to-end encryption, BAA, EHR integration.
Reviews: Capterra 5/5 ★, G2 4.8/5 ★ (“best platform for clinics, responsive support”).
URL: emitrr.com/hipaa-compliant-chatbot | emitrr.com/chatbot-for-healthcare

2. BastionGPT – Secure AI Assistant for Documentation

Use: Clinical documentation, scribing, summarization.
HIPAA Features: BAA included, data never shared externally, encryption at rest & transit.
Reviews: 4.8/5 ★ on SoftwareAdvice & Capterra (“like ChatGPT but private, saves hours”).
URL: bastiongpt.com | bastiongpt.com/security

3. TigerConnect – Secure Clinical Communication

Use: HIPAA-secure team messaging, alerts, scheduling, with chatbot extensions.
HIPAA Features: Enterprise-grade hospital communication system, secure workflows.
Reviews: Widely adopted in hospital systems; praised for reliability and compliance.
URL: tigerconnect.com

Faculty Research Focus: MIT Sloan & J-Clinic

Regina Barzilay
As Distinguished Professor for AI and Health, Barzilay applies natural language processing (NLP) and deep learning to oncology. Her models leverage unstructured patient records and imaging to improve risk stratification, optimize chemotherapy regimens, and advance precision medicine in breast cancer.

Dimitris Bertsimas
Bertsimas integrates operations research, predictive analytics, and optimization algorithms to address systemic health care challenges. His research translates large-scale data sets into decision-support tools for hospital resource allocation, disease progression modeling, and public health policy, ensuring scalable and cost-effective solutions.

Tommi Jaakkola
Jaakkola’s work spans machine learning theory, probabilistic modeling, and representation learning. In health contexts, his contributions enable robust inference from noisy data, powering innovations in genomic analysis, drug discovery, and interpretable AI systems that balance predictive accuracy with transparency.

Dina Katabi
Katabi, a MacArthur Fellow, pioneers wireless sensing and computational imaging to extract clinical signals unobtrusively. By applying signal processing and AI inference, her team transforms ambient wireless data into continuous monitoring tools for Parkinson’s disease, sleep disorders, and cardiac health, expanding care beyond hospital walls.

David Sontag
Sontag develops causal inference frameworks and probabilistic graphical models for electronic health record (EHR) analysis. His research emphasizes overcoming bias and missing data to enhance disease prediction, treatment effect estimation, and personalized care pathways at the population level.

Collin Stultz
Stultz combines expertise in computational biology and cardiology, applying biomedical signal processing and Bayesian modeling to predict adverse cardiac events. His work highlights the synergy between mechanistic models and data-driven AI, advancing reliability in ICU mortality prediction and patient care optimization.

As a graduate of Stanford’s Artificial Intelligence in Healthcare certificate program, I explored how machine learning (ML), natural language processing (NLP), and predictive analytics are reshaping clinical care. The course emphasized supervised and unsupervised learning, causal inference, bias mitigation, and model interpretability, all applied to electronic health records (EHRs), radiology imaging, and biomedical signals.

The program highlighted how AI can be ethically integrated into learning health systems, improving patient safety, care quality, and population-level outcomes. Case studies ranged from automated radiology reporting to ICU mortality prediction, demonstrating both the promise and limitations of AI in practice.

While the curriculum was not designed as a coding bootcamp, it prepared me to pursue further studies in the technical health AI tech stack—from Python, TensorFlow, and PyTorch for model development, to FHIR APIs, SQL-based EHR queries, and cloud platforms (AWS, GCP) for deployment—should I choose to deepen my engineering expertise.

Ultimately, this program gave me a rigorous foundation in applying AI responsibly and safely in healthcare, and positioned me to collaborate effectively across medicine, data science, and policy as this field continues to evolve.

Matthew P. Lungren https://profiles.stanford.edu/matthew-lungren

Imaging AI, deployment science https://pubmed.ncbi.nlm.nih.gov/37224199/

Serena Yeung-Levy https://profiles.stanford.edu/serena-yeung

Computer vision, clinical video analytics https://pmc.ncbi.nlm.nih.gov/articles/PMC6550251/

Tina Hernandez-Boussard https://profiles.stanford.edu/tina-hernandez-boussard

Responsible AI, evaluation & fairness https://pubmed.ncbi.nlm.nih.gov/32594179/

Laurence C. Baker https://profiles.stanford.edu/laurence-baker

Health economics, systems & incentives https://pubmed.ncbi.nlm.nih.gov/27503970/

Steven C. Bagley https://profiles.stanford.edu/steven-bagley

Clinical data methods, informatics pedagogy https://pubmed.ncbi.nlm.nih.gov/11576808/

Nigam H. Shah https://profiles.stanford.edu/nigam-shah

Learning health systems, EHR mining, causal inference https://pubmed.ncbi.nlm.nih.gov/34423364/

David Magnus https://profiles.stanford.edu/david-magnus

Bioethics, regulation, ambient intelligence https://pubmed.ncbi.nlm.nih.gov/33358138/

Curtis P. Langlotz https://profiles.stanford.edu/curtis-langlotz

Imaging informatics, AI governance & translation https://pubmed.ncbi.nlm.nih.gov/39802660/

Deep Learning & COVID X-Ray Detection

📚 Lesson: Deep Learning & COVID X-Ray Detection

1. How Images Are Represented for AI

When computers "see" an image, they don’t see a photograph the way humans do. Instead, each image is translated into numbers representing pixel intensity.

A grayscale image is stored as a 2D grid (matrix) of numbers, each showing brightness (0 = black, 255 = white).
A color image (RGB) is stored as 3 grids (layers) — one for red, one for green, and one for blue.

👉 This layered grid representation is essential because it preserves spatial structure (shapes, edges, textures).
❌ Flattening an image into a single line of numbers loses that spatial relationship and weakens the model’s ability to detect patterns.

2. Neural Network Architectures for Images

Not all neural networks are created equal. Different architectures excel at different data types:

MLPs (Multi-layer Perceptrons): Good for tabular data, but inefficient for images.
RNNs (Recurrent Neural Networks): Designed for sequences (text, speech, time series).
GANs (Generative Adversarial Networks): Best at creating or enhancing images, not classifying them.
CNNs (Convolutional Neural Networks): ⭐ The workhorse for medical imaging. They slide filters across an image to detect edges, shapes, and patterns, layer by layer.
👉 For X-ray analysis, CNNs are the standard.

3. Types of Prediction Questions

Machine learning problems can be grouped into types:

Regression → Predicts a continuous number (e.g., predicting blood glucose from features).
Multi-label classification → Assigns multiple possible labels (e.g., pneumonia + COPD on the same X-ray).
Binary classification → A simple yes/no decision (COVID-positive vs COVID-negative).
Sequence-to-sequence translation → Maps one sequence to another (e.g., language translation, speech-to-text).

👉 COVID detectors are binary classifiers because they decide: “Does this patient have COVID-related findings? Yes or No.”

4. Boosting Performance with Extra Data

The performance of an AI system depends on both the architecture and the training data. Hospitals can contribute resources to make detectors stronger:

Text reports: Radiology notes paired with images allow models to learn from both visual and language data.
Segmentation labels: Marking lung regions in even a small subset of images helps the AI focus on what matters.
Existing COVID detectors: Using pretrained models accelerates learning (transfer learning).
Large datasets of chest X-rays (even non-COVID): More data improves generalization and helps the model learn normal vs abnormal anatomy.

👉 In real-world medical AI, all of these actions are beneficial.

💡 Key Takeaways Before the Test

Images → represented as layered grids of pixels.
CNNs → the go-to network for image classification.
COVID detection → a binary classification problem.
More resources (text, labels, pretrained models, large datasets) → stronger performance.

Data for Research with AI

1. Data Representation

EHR is structured data (labs, vitals, demographics, comorbidities).
Represented as a 1D feature vector: each element = one clinical measurement.
❌ Not represented as pixel intensities or image grids (those are for radiology).
✅ Example: [Age, Temp, Ferritin, WBC, O₂ saturation].

2. Framing ML Tasks in Healthcare

Regression: Predict continuous outcomes.
- Example: Days until intubation (for patients who eventually need it).
Binary classification: Predict Yes/No outcomes.
- Example: Will this patient be admitted? (not the focus here)
Multiclass classification: Predict categorical ranges.
- Example: Intubation within 0–4, 5–9, 10–14, or never.
✅ Task framing must align with clinical problem statement.

3. Feature Selection & Data Leakage

Valid features: Labs (Ferritin, D-Dimer, WBC), vitals, demographics (age).
Invalid features (leak future info or PHI):
- Invasive mechanical ventilation date (outcome itself).
- Ventilator settings (only known post-intubation).
- Exact birth date or admission date (should be transformed into age or relative time).
✅ Always ask: Would this information be available at prediction time?

4. Data Quality & Healthcare Workflow Errors

Clinical datasets are messy; possible issues include:

Labs logged after intubation → gives model access to future info.
Immediate ED intubation → no labs collected, missing data.
Transfers from other systems → incomplete history.
Multiple visits → less of a concern if separated by encounter IDs.

5. Key Takeaways Before the Quiz

EHR = 1D feature vectors, not images.
Choose ML framing carefully: regression or multiclass depending on timeline prediction.
Exclude features that leak future outcomes.
Anticipate and mitigate data errors and missingness.
Always align models with the clinical question of interest.

Page updated

Google Sites

Report abuse