Skip to content

AI × HEALTHCARE

Explore AI in Healthcare

A brain signal becomes a voice. A protein sequence becomes a research clue. A referral moves closer to its next step. Explore how different kinds of AI contribute to healthcare, what the evidence shows, and where you could begin learning.

Follow eight visual examples, then choose a research question you could develop into a project.

Watch the examplesExplore the program
BRAIN SIGNAL → DECODER → VOICE
From attempted speech to a voiceConceptual diagram of an implanted recording array, a trained neural decoder and a sound waveform. These are schematic signals, not a neural recording or arbitrary thoughts.
From attempted speech to sound.Conceptual illustration · Explore the evidence below.
Published by STEMtiq AISources checked October 25–29, 202619 min read

THE BASICS

What medical question is the system answering?

Healthcare AI includes several kinds of systems. Computer vision can classify a pattern in an image. Prediction models estimate an outcome. Language systems can rank papers or generate responses in a consultation. An agent connects tools across several steps, such as retrieving papers and organizing their findings. Each application needs an evaluation suited to its task.[1][3][4]

Ask which decision the output supports. A search result helps a reader locate a paper; assessing that paper requires reading its methods and findings. An image model identifies a pattern; its clinical meaning depends on the intended use and surrounding care process.

AI AT WORK

Eight ways AI contributes to healthcare

Watch the information move through one application at a time. Each illustration explains a process. The evidence beside it shows what was actually studied or documented.

Explore eight illustrated examples and the evidence behind each one.

Neuroscience · Communication

Turning attempted speech into a voice

UC Berkeley / UCSF

From attempted speech to a voiceConceptual diagram of an implanted recording array, a trained neural decoder and a sound waveform. These are schematic signals, not a neural recording or arbitrary thoughts.
  1. 01Brain signal
  2. 02Decoder
  3. 03Voice
Generate a voice.The system streams audible speech. This is personalized speech decoding, not arbitrary thought reading.

Conceptual illustration. Schematic brain signals, not a neural recording.

A trained neural decoder maps speech-related brain activity to audible speech in a participant’s personalized voice.

Single-participant research

2025: 47.5 words per minute on the full-vocabulary task, in one woman with severe paralysis.

It explores a way for people who cannot speak to communicate through their own attempted speech.

Think of a translator learning the relationship between this person’s neural signals and the speech she is trying to produce.
Sources[6] Nature Neuroscience[7] National Institutes of Health
Evidence limits and what you could learn

The workflow

Attempted speech → Implanted recording → Neural decoder → Streaming voice.

  1. Record the signal. An implanted array records activity while the participant attempts to speak.
  2. Decode the pattern. A trained decoder maps this participant’s speech-related signals to speech output.
  3. Generate a voice. The system streams audible speech. This is personalized speech decoding, not arbitrary thought reading.

Human check

Researchers train and evaluate the decoder for the participant and assess its errors and limitations.

Evidence limits

Experimental work in one participant. It requires an implant and training, and does not read arbitrary thoughts. The reported 80-millisecond increments describe audio output chunks, not total system latency.

One term: Neural decoder

A model that maps recorded neural activity to a defined output, such as speech.

Learning foundations

Signals and time series · Machine learning · Error measurement

A possible student starting point

Build a cited brain-research explorer that compares what speech-decoding studies tested, achieved and could not establish.

Biology · Protein structure

Predicting a protein’s three-dimensional shape

AlphaFold · Google DeepMind / EMBL-EBI

From a protein sequence to a predicted shapeConceptual amino-acid beads move into a schematic folded chain. A highlighted region illustrates a confidence check. This is not an actual AlphaFold structure.
  1. 01Sequence
  2. 02Prediction
  3. 03Confidence
Inspect confidence.Researchers examine confidence and use experiments to investigate the predicted structure and its function.

Conceptual illustration. Schematic fold, not an actual AlphaFold model.

Predicts a protein’s structure from its amino-acid sequence, giving researchers a starting point for investigating how it works.

Research database

The 2022 database expansion released over 200 million predicted protein structures.

A protein’s shape helps researchers ask how it might interact with other molecules and what experiments to try next.

The sequence is a chain of building blocks. The folded shape helps us investigate what that chain might do.
Sources[8] EMBL-EBI[9] Google DeepMind
Evidence limits and what you could learn

The workflow

Protein sequence → Structure prediction → Confidence check → Research and experiments.

  1. Start with a sequence. A protein begins as a sequence of amino acids. The sequence is the model’s input.
  2. Predict the fold. The model predicts how that chain may arrange itself in three dimensions. This drawing is schematic.
  3. Inspect confidence. Researchers examine confidence and use experiments to investigate the predicted structure and its function.

Human check

Researchers inspect confidence and use additional evidence or experiments to investigate the predicted structure.

Evidence limits

These are computational predictions with varying confidence. A predicted structure alone does not establish a protein’s function, a useful drug or a treatment benefit. The count describes the 2022 expansion.

One term: Amino acid

One of the building blocks joined together in a protein sequence.

Learning foundations

Protein biology · Model predictions · Confidence and validation

A possible student starting point

Create a protein explorer that retrieves a public structure record and explains its confidence and research context.

Genetics · DNA variants

Finding DNA changes worth investigating

AlphaMissense · Google DeepMind

Investigating a missense variantA highlighted DNA rung is connected to a protein residue and three interpretation categories: likely benign, uncertain and likely pathogenic. No real variant is scored.
  1. 01DNA change
  2. 02Protein context
  3. 03Interpretation
Keep uncertainty visible.Predictions can be likely benign, uncertain or likely pathogenic. They guide investigation and do not establish a diagnosis.

Conceptual illustration. Illustrative DNA. No real variant is scored.

Predicts whether a missense variant, a DNA change that alters one amino acid in a protein, is more likely harmful or benign.

Computational predictions

2023: predictions for 71 million possible human missense variants; 89% categorized as likely benign or likely pathogenic.

The predictions can help researchers prioritize variants for closer investigation alongside other biological and clinical evidence.

A prediction helps decide where to look next. We still need evidence to decide what the change means.
Sources[10] Science[11] Google DeepMind
Evidence limits and what you could learn

The workflow

DNA variant → Protein context → Predicted effect → Expert evidence review.

  1. Spot a sequence change. A missense DNA variant changes one amino acid in the protein it encodes.
  2. Consider protein context. The model considers sequence and structural context to predict a possible effect.
  3. Keep uncertainty visible. Predictions can be likely benign, uncertain or likely pathogenic. They guide investigation and do not establish a diagnosis.

Human check

Qualified researchers and clinicians interpret predictions alongside other evidence; a score alone does not establish a diagnosis.

Evidence limits

89% is the share assigned to those categories, not diagnostic accuracy. The remaining variants were uncertain. Predictions are not 71 million clinically confirmed findings and cannot establish a diagnosis on their own.

One term: Missense variant

A DNA change that changes one amino acid in a protein.

Learning foundations

DNA-to-protein basics · Biological databases · Probability and uncertainty

A possible student starting point

Compare public research about one variant, keeping model predictions separate from experimentally established findings. Use no personal DNA.

Drug discovery · Lung disease

Taking an AI-designed drug candidate into a trial

Rentosertib · Insilico Medicine

From a biological target to clinical testingA schematic protein pocket is connected to a generic candidate molecule and a clinical-trial record. The drawing is not the structure of rentosertib and shows no treatment result.
  1. 01Target
  2. 02Candidate
  3. 03Trial
Test in clinical trials.Clinical trials study safety and possible benefit in people. Reaching a trial does not establish that a treatment works.

Conceptual illustration. Schematic molecule, not the structure of rentosertib.

AI helped prioritize the protein TNIK as a target and design a small-molecule candidate for idiopathic pulmonary fibrosis, a disease that scars the lungs.

Published early clinical trial

A 2025 randomized phase 2a trial studied 71 people for 12 weeks, with safety as its primary endpoint.

It is a concrete example of AI contributing to both target selection and molecule design, followed by testing in people.

AI can propose a promising molecule. Biology still gets the final vote.
Sources[12] Nature Medicine[13] Insilico Medicine, via PR Newswire
Evidence limits and what you could learn

The workflow

Disease biology → Target and molecule design → Laboratory tests → Clinical trials.

  1. Identify a biological target. AI helped prioritize TNIK, a protein investigated as a target in lung fibrosis.
  2. Propose and test molecules. AI helped design a small-molecule candidate. Laboratory tests must investigate its biological properties.
  3. Test in clinical trials. Clinical trials study safety and possible benefit in people. Reaching a trial does not establish that a treatment works.

Human check

Laboratory scientists and clinical researchers test safety and possible benefit. A proposed molecule is only a candidate.

Evidence limits

The small, short 2025 trial does not establish long-term treatment benefit. Lung-function findings were secondary endpoints. The company announced phase III initiation in July 2026; the early trial should not be mistaken for its current development stage or proof of approval.

One term: Biological target

A molecule or biological process that an intervention is intended to affect.

Learning foundations

Biological data and APIs · Molecular representations · Clinical trial interpretation

A possible student starting point

Build a target-evidence agent that connects a protein’s biological role, published studies and trial outcomes, like the Neuro Discovery Agent.

Medical imaging · Eye screening

Screening retinal photographs for diabetic eye disease

IDx-DR, now LumineticsCore · Digital Diagnostics

Retinal photographs, image quality and follow-upAn illustrated retina connects to an image-quality question and a follow-up icon. This is not a patient scan, device analysis or screening result.
  1. 01Photograph
  2. 02Quality
  3. 03Follow-up
Connect screening to follow-up.The result guides the next screening or referral step within the device’s specific indication. Care continues with the healthcare team.

Conceptual illustration. Illustrated retina, not a patient image or screening result.

Checks retinal image quality and analyzes the photographs for a defined level of diabetic retinopathy.

FDA-authorized screening device

FDA authorized IDx-DR in 2018 after review of a study enrolling 900 people at 10 primary care sites.

It brings a defined eye-screening task into a primary care workflow, with results that inform referral and follow-up.

A useful healthcare model has to recognize when the photograph is too poor to judge.
Sources[14] U.S. Food and Drug Administration[15] Digital Diagnostics
Evidence limits and what you could learn

The workflow

Retinal photos → Image-quality check → Screening result → Referral or follow-up.

  1. Start with a retinal photograph. The system receives an image from a compatible retinal camera. This drawing shows the idea, not a patient scan.
  2. Check whether the image can be used. Image quality matters. An unusable photograph cannot support a reliable screening result.
  3. Connect screening to follow-up. The result guides the next screening or referral step within the device’s specific indication. Care continues with the healthcare team.

Human check

Healthcare staff follow the authorized indication, image-quality requirements and appropriate referral or follow-up process.

Evidence limits

The original authorization was for specified equipment and adults with diabetes without a previous diabetic-retinopathy diagnosis. It does not assess every eye disease or provide treatment. FDA reported 87.4% sensitivity and 89.5% specificity in 819 analyzable participants, not all 900 enrolled. These are different measures, not one accuracy score.

One term: Retina

The light-sensitive tissue at the back of the eye.

Learning foundations

Computer vision · Sensitivity and specificity · Quality checks

A possible student starting point

Build an educational comparison of published screening results and image-quality rules, rather than a clinical screening tool.

Biotech · Protein designYC S20

Designing proteins for a specific job

Nabla Bio

Protein designs need laboratory measurementsA schematic target connects to three antibody candidates and laboratory assay wells. No binding result or laboratory measurement is shown.
  1. 01Target
  2. 02Designs
  3. 03Lab tests
Measure what the designs actually do.Laboratory assays test properties such as binding and biological function. Experiments provide evidence that a computer prediction cannot supply alone.

Conceptual illustration. Schematic designs. No laboratory measurements are shown.

Uses generative AI to propose antibody and other protein designs, with biological experiments measuring properties such as binding and function.

Company-described research platform

Its official platform describes AI protein design paired with laboratory measurements. YC confirms the Summer 2020 batch.

It connects generating a possible design with measuring whether that design has useful biological properties.

AlphaFold asks what shape a protein may take. Protein design asks what sequence we might create to perform a particular job.
Sources[16] Nabla Bio[17] Y Combinator
Evidence limits and what you could learn

The workflow

Biological target → Generated protein designs → Laboratory measurements → Evaluate properties.

  1. Choose a biological target. Researchers define the protein and the properties they want an antibody to have.
  2. Propose candidate designs. A generative model proposes protein sequences. A plausible design is a candidate to investigate, not a finished medicine.
  3. Measure what the designs actually do. Laboratory assays test properties such as binding and biological function. Experiments provide evidence that a computer prediction cannot supply alone.

Human check

Laboratory researchers measure the properties of proposed designs and decide which candidates deserve further investigation.

Evidence limits

Platform descriptions establish what the company says its system does, not independently demonstrated patient benefit. Early drug research is different from an approved therapy. Nabla Bio is a different company from Nabla, the clinical documentation product.

One term: Antibody

A protein that can recognize and bind a particular target.

Learning foundations

Protein sequences · Generative models · Experimental validation

A possible student starting point

Create a public protein-research explorer that compares target evidence and experimental findings. Reproducing a drug-design platform requires much deeper expertise and laboratory work.

Healthtech · Care coordinationYC W23

Moving a referral beyond the paperwork

Tennr

Checking a fictional referralA clearly fictional document has a missing referral reason. Structured fields connect to a draft request and staff review. No patient information is used and no message is sent.
  1. 01Referral
  2. 02Missing detail
  3. 03Review
Prepare the next action.A request for the missing detail can be drafted for review. This teaching example sends no message and uses no patient data.

Conceptual illustration. Fictional referral. No message is sent.

Reads referral documents, extracts information, identifies missing requirements and coordinates the next operational step.

Company-described workflow

Official product and YC descriptions cover document processing and referral coordination. YC confirms the Winter 2023 batch.

Getting care also depends on documents, requirements and communication reaching the right people. This is a useful bridge from AI predictions to agent workflows.

A chatbot might explain what is missing. A connected workflow can identify it, request it and track the next step.
Sources[18] Tennr[19] Y Combinator
Evidence limits and what you could learn

The workflow

Incoming referral → Extract fields → Request missing information → Route and track.

  1. Read an incoming referral. A referral may arrive as a fax or PDF. The workflow first has to turn the document into usable information.
  2. Identify what is missing. Extracted fields are checked against the workflow’s requirements. This fictional form is missing a reason for the referral.
  3. Prepare the next action. A request for the missing detail can be drafted for review. This teaching example sends no message and uses no patient data.

Human check

Authorized staff review exceptions and uncertain fields, using approved workflow rules before real routing or communication.

Evidence limits

Product capabilities and company-reported volumes do not independently prove fewer errors, faster treatment or better patient outcomes. An educational version should send uncertain fields to a person for review.

One term: Structured data

Information organized into named fields that software can check.

Learning foundations

Document extraction · Structured data and APIs · Workflow logic and review

A possible student starting point

Build a referral-checklist agent using fictional documents. Flag missing fields and require human approval before any message or routing action.

Clinical knowledge · Evidence retrievalYC W23

Bringing medical sources into the answer

Glass Health

Connecting a question to inspectable sourcesA question connects to research documents and a cited draft for clinician review. The illustration generates no clinical answer and contains no fabricated paper titles.
  1. 01Question
  2. 02Sources
  3. 03Cited draft
Make the draft inspectable.Citations help a clinician inspect the supporting material. A citation alone does not prove that a conclusion is correct.

Conceptual illustration. Illustrated retrieval. No clinical answer is generated.

Supports clinicians with literature search, clinical questions and draft reasoning or plans, with source citations.

Company-described clinical support

Developer documentation describes literature search and cited answers. YC confirms the Winter 2023 batch.

The useful connection is between a question, relevant medical literature and an answer whose sources a clinician can inspect.

Think of a research assistant that brings relevant sources to the clinician’s desk. The sources and the reasoning still need checking.
Sources[20] Glass Health[21] Y Combinator
Evidence limits and what you could learn

The workflow

Clinical question → Retrieve medical sources → Draft with citations → Clinician review.

  1. Start with a clinical question. The question defines what information needs to be found. This scene does not contain a patient case or a medical answer.
  2. Retrieve relevant medical sources. The system searches medical literature and guidance so its draft can be connected to identifiable sources.
  3. Make the draft inspectable. Citations help a clinician inspect the supporting material. A citation alone does not prove that a conclusion is correct.

Human check

Clinicians inspect the cited material, its relevance and the draft reasoning before using it in professional work.

Evidence limits

Citations alone do not establish that a conclusion is correct or that using the product improves patient outcomes. This is support for professional review, not evidence of autonomous diagnosis. YC membership is not clinical validation.

One term: Retrieval

Finding relevant information from a defined collection of sources.

Learning foundations

Retrieval and tool use · Citation checks · Evaluation and uncertainty

A possible student starting point

Build an educational literature agent for one neuroscience question, showing the sources and where its conclusion stops.

The examples identify published research, computational predictions, a research database, a regulatory authorization or a company-described capability. A YC label identifies an accelerator batch; it is not clinical validation or an endorsement.

Go deeper: three notes on evaluating healthcare AI

Looking for patterns in medical images

Regulated device applications and published validation research

A clinician needs information from a medical image for a specified clinical task.

The material
Images such as retinal photographs or chest X-rays, with a documented collection setting and reference labels used for evaluation.
AI’s role
An image model can classify a pattern or provide information for a defined clinical task. The FDA lists authorized AI-enabled device examples, including retinal-image analysis.
The human decision
Read the device’s intended use: which task it supports, who receives the output, and how that information enters the care process.

The limitZech and colleagues found that chest X-ray models did not consistently retain their performance across hospitals. Device authorization also applies to a particular intended use; it does not validate an unrelated chatbot.[1][2]

Finding evidence in biomedical literature

An established research-search service; proposed student evidence workflows

A research question can lead to many papers that study different populations, methods, and outcomes.

The material
Publication records, abstracts and accessible full-text papers.
AI’s role
PubMed’s Best Match ordering uses machine learning to rank relevant citations. A separate language-assisted workflow could help organize selected findings into an evidence table.
The human decision
A reader checks the original paper, verifies each extracted statement, and decides whether it actually answers the question.

The limitA highly ranked paper still needs critical reading. Follow the full-text link to check methods and limitations that may be absent from the abstract.[3]

Studying how an AI gathers clinical information

Published research using simulated consultations

Answering a written medical question does not show how well a system can ask useful follow-up questions.

The material
In the AMIE study, patient-actors used prepared scenarios in text conversations.
AI’s role
Researchers investigated a language system designed for diagnostic dialogue, assessing several aspects of its responses.
The human decision
Specialist physicians and patient-actors evaluated the consultations, including the diagnostic responses and qualities of the conversation.

The limitThe study’s text-chat setting was unfamiliar to the participating physicians. It was not a test of ordinary clinical care or evidence of improved patient health.[4]

CASE STUDY

How AMIE’s diagnostic-dialogue study was evaluated

The researchers compared text consultations under controlled, simulated conditions.

Tu and colleagues compared AMIE with primary care physicians in a randomized crossover study. Patient-actors held text consultations with each about the same prepared case.[4]

Specialists and actors evaluated diagnostic responses and conversation qualities. The comparison addressed a simulation, rather than routine care or patient-health outcomes.[4]

When reading a medical AI study, identify the input, comparison and measured outcome. Then check whether the headline describes the same setting. Evidence from a simulated consultation leaves questions about routine care and patient outcomes unanswered.

Follow the process Open a step to see what happens.

  1. A defined setting

    Prepared scenarios and patient-actors formed a simulated text consultation.

  2. A comparison

    The same case was discussed with AMIE and with a physician.

  3. An evaluation

    Specialists and patient-actors assessed the resulting consultations.

  4. An open question

    Routine clinical use and patient benefit need further evidence.

A simplified account of the cited case. Read the original source for its full methods.

Describe the result as evidence about simulated text consultations. Routine clinical use and patient benefit require additional evidence.[4]

FROM APPLICATION TO AGENT

What changes when AI connects several steps?

A model can predict a structure, recognize a pattern or draft text. An agent can select and connect tools around a goal. For a research question, that might mean finding the right biological record, retrieving evidence, checking a trial and producing a brief with sources. The tools and the checks need to be visible enough to inspect.

THE WORKFLOWReviewing biomedical evidence

Select a step to see what happens.

Research paper

Read what was actually studied

Identify the research question, population, comparison, and measured outcome in the original paper.

Evidence table

Organize the information

A language-assisted workflow can propose entries in an evidence table. Each entry still needs a passage, table, or figure behind it.

Source check

Keep the conclusion within the evidence

Check each entry against its source. A simulated consultation result does not establish a benefit in routine patient care.

AI in Healthcare: applications

Illustrative evidence workflow. It does not make a clinical decision.

A useful first build can be a Neuro Discovery Agent: an educational tool that investigates one question in public neuroscience research, keeps findings linked to their sources, and distinguishes a biological clue from evidence that a treatment helps people.

SKILLS & STUDY

What to learn first

Scientific reading
Identify the question, study design, participants, comparison, and outcome. Learn enough biology to explain what the paper measured.
Data reasoning
Ask where examples came from, how labels were assigned, and whether evaluation data represent the proposed use. Separate a model’s score from its consequences.
Verification and communication
Keep citations beside working notes. Explain disagreement or missing information clearly, and distinguish the authors’ findings from your interpretation.
Technical foundations
A reading project can begin with a spreadsheet. Automated extraction needs programming and output checks. Training an image model adds statistics, dataset preparation, and evaluation work.

Follow an interest into a learning direction

You do not need to reproduce a specialist system to begin. Learn to define a question, work with the right information, test an output and explain what it does and does not establish.

If you were drawn to…Foundations to investigateA reasonable starting scope
Brain signals and speechSignals, model inputs/outputs, evaluationCompare published speech-decoding evidence.
Proteins, genetics or drug discoveryBiology, biological identifiers, data retrieval, uncertaintyBuild a public target/protein evidence explorer.
Medical imagesLabeled data, image quality, sensitivity and specificityExplain published evaluations or create a teaching dashboard.
Referrals and care coordinationStructured fields, tool calls, workflow state, reviewUse fictional referrals and simulated actions.
Medical literatureRetrieval, citations, structured extraction, checksCreate a brief from a narrow, mentor-selected source set.

PROJECT IDEAS

Two ways to investigate biomedical evidence

Choose a question, inspect the requirements, and keep a brief for later. Each proposal includes a way to check the work.

Save an idea in this browser or copy its brief to discuss with a mentor.

No account needed
PROJECT IDEA 01Proposed project

Map evidence for one biomedical question

How do selected published studies differ in what they investigate and measure?

Intended output
A citation-linked table and short research brief comparing methods, reported findings, and unanswered questions.
Skills, sources and checks
Starting knowledge
Basic biology, careful reading, and spreadsheet skills. Automated extraction would add programming and verification requirements.
Material to consider
A mentor-selected set of public abstracts and accessible papers on one narrow question, such as a specific CRISPR research approach. Record why each paper was included.
How to judge the result
Check every table entry against its source. Distinguish study types and mark information unavailable in the selected material.

Keep conclusions limited to the selected studies and the research question you defined.

PROJECT IDEA 02Proposed project

Check whether a public-health claim matches its source

Does a public explanation describe what the cited study actually found?

Intended output
An annotated claim-to-source comparison showing supported wording, missing context, and a proposed clearer explanation.
Skills, sources and checks
Starting knowledge
Identify a study’s population, comparison, and outcome. Understand the difference between an association and evidence of causation.
Material to consider
A small collection of public institutional research summaries and their openly accessible underlying studies; no patient information or personal health questions.
How to judge the result
Trace each statement to a passage, table, or figure. Have a knowledgeable reviewer check the interpretation and the limits of the proposed rewrite.

A science-writing project examining whether a public claim matches its evidence.

These project ideas use public research or synthetic material. They do not involve patient information, diagnosis, treatment recommendations, or claims of clinical readiness.

MENTOR SUPPORT

Turn an interest into a project you can explain

The Agentic AI Scholars Program gives students a guided route into building and evaluating an agent. A healthcare direction could focus on public research, structured evidence and citation checks, with a mentor helping you keep the question manageable. Explore the curriculum, current schedule, support and tuition to decide whether it fits your interest.

5 weeks · 25 live hours · Grades 9-12

Explore Agentic AI ScholarsSTEMtiq Agentic AI Scholars ProgramOne shared agent-building program. Discuss project fit, current dates, tuition, and support on the program page.

COMMON QUESTIONS

What students ask

Can I explore healthcare AI before taking advanced biology?

Yes. Start by identifying what an application does and how it was evaluated. Choose papers whose basic terminology you can explain. Specialized study interpretation and model training require more biology, statistics and guidance.

Would my first project need medical images or patient records?

No. An evidence table built from selected public papers can be a useful first project. The ideas in this guide do not require patient information or ask you to interpret someone’s scan.

How much coding would an evidence project require?

You can compare a few papers with a document and spreadsheet. Automating retrieval or extraction adds programming and checks for missing, incorrect or unsupported statements. Decide which version of the project you can evaluate yourself.

Does an authorized AI device prove that medical AI is reliable in general?

No. The FDA evaluates devices through applicable pathways for their intended uses. That is different from validating all AI tools, a general-purpose chatbot, or an unrelated use of the same technology.[1]

Is pre-med the major I would need to choose?

Pre-med is a preparation path, not a required major. Students can study different subjects while completing medical-school prerequisites. Check individual schools’ requirements with a prehealth advisor.[5]

Are these animations running real medical AI?

No. They are original diagrams explaining a process. Published findings and product documentation are identified beside each example. The animations do not process patient information or generate a medical conclusion.

Could my first project reproduce one of these systems?

A first project should have a smaller, testable scope, such as retrieving research, organizing evidence or checking fictional documents. Specialist clinical and laboratory systems require substantially more domain knowledge, testing and infrastructure.

Plan a source-backed research brief

Choose a biomedical question, identify the papers you would use, and decide how to check the resulting explanation. The program page describes mentor support for developing this kind of AI-agent project.

Explore Agentic AI ScholarsSTEMtiq Agentic AI Scholars Program

Sources & further reading

  1. [1]
  2. [2]
  3. [3]
    National Library of Medicine: PubMed User Guide (2026-09-30) · Source checked October 25, 2026Return to guide ↑
  4. [4]
  5. [5]
  6. [6]
  7. [7]
  8. [8]
  9. [9]
  10. [10]
  11. [11]
  12. [12]
  13. [13]
  14. [14]
  15. [15]
  16. [16]
    Nabla Bio: Platform (Undated) · Source checked October 29, 2026Return to guide ↑
  17. [17]
    Y Combinator: Nabla Bio (Undated) · Source checked October 29, 2026Return to guide ↑
  18. [18]
    Tennr: Product (Undated) · Source checked October 29, 2026Return to guide ↑
  19. [19]
    Y Combinator: Tennr (Undated) · Source checked October 29, 2026Return to guide ↑
  20. [20]
    Glass Health: API documentation (Undated) · Source checked October 29, 2026Return to guide ↑
  21. [21]
    Y Combinator: Glass Health (Undated) · Source checked October 29, 2026Return to guide ↑

Sources inform this guide; their organizations do not endorse STEMtiq. Project ideas and diagrams are original editorial explanations.