HCC 3030 · Week 6
AI inHealthcare A thousand places to help, from the first signal of disease to the discovery of a new treatment.
CARE
HOME CLINIC LAB HOSPITAL PUBLIC HEALTH DISCOVERY
This lecture is deliberately opportunity-led. Healthcare is not a single task, institution, or user. It is a connected system of prevention, diagnosis, treatment, operations, recovery, public health, and scientific discovery.
Students should leave able to identify many forms of healthcare AI, connect each to the right technical method and workflow, and distinguish a promising model from a clinically valuable system.
[Sources]
Start with the landscape
Healthcare is not one place.
kitchen table pharmacy school clinic ambulance primary care radiology suite laboratory operating room hospital ward rehabilitation gym public-health office research bench
Every setting has different data, decisions, experts, time pressure, and definitions of success.
Ask students to picture healthcare as an ecosystem rather than a visit with a physician. Many of the highest-value opportunities happen before the appointment, between visits, after discharge, or far upstream in science and operations.
The WHO describes uses across diagnosis, clinical care, drug development, disease surveillance, outbreak response, and health-system management.
[Sources]
A single journey
One person’s care journey contains dozens of moments where AI could help.
1 NOTICE A wearable detects a change.
→ 2 ACCESS Triage finds the right setting.
→ 3 DIAGNOSE Images, labs, and history align.
→ 4 TREAT A plan fits the person.
→ 5 RECOVER Home monitoring catches setbacks.
schedule document coordinate educate bill learn
Use the journey to surface both visible clinical work and invisible coordination work. An algorithm can detect a signal, but the system must also schedule, communicate, document, route, and follow up.
The lower row is important. Removing friction can create as much value as improving a diagnostic benchmark because it changes capacity and continuity.
[Sources]
The central claim
The promise is not replacing medicine. It is changing what medicine can notice, reach, and accomplish.
NOTICE weak signals earlier
REACH people outside specialist centers
PERSONALIZE care to a changing patient
COORDINATE complex work across teams
DISCOVER new biology and treatments
Frame AI as an expansion of sensing, memory, pattern recognition, simulation, coordination, and scientific search. Some uses automate tasks, but many uses create capabilities that were previously too slow, expensive, or geographically scarce.
The key question is not only whether AI can perform a task. It is what new form of care becomes possible when the task becomes faster, cheaper, continuous, or widely available.
[Sources]
Why the opportunity is so large
Healthcare is constrained by both scarce expertise and enormous system cost.
11 million projected global health-worker shortfall by 2030
$5.3 trillion U.S. health spending in 2024
18.0% of U.S. gross domestic product in 2024
AI matters economically when it improves outcomes, expands capacity, or removes work that consumes expertise without helping patients.
The WHO currently projects a global shortfall of 11 million health workers by 2030, concentrated in low- and lower-middle-income countries. CMS reports that U.S. health spending reached $5.3 trillion in 2024, equal to 18 percent of GDP.
Do not treat these numbers as proof that any AI product will save money. They explain why tools that expand capacity, reduce avoidable work, or prevent expensive downstream harm have unusually large potential.
[Sources]
This is already a deployment story
Predictive AI is already embedded in most U.S. hospitals.
71%
of surveyed non-federal acute-care hospitals reported predictive AI integrated with the EHR in 2024. Common uses included inpatient risk, readmission, early disease detection, scheduling, billing, and treatment recommendations.
Small, rural, independent, and critical-access hospitals reported lower adoption.
The federal data brief reports 71 percent adoption in 2024, up from 66 percent in 2023. It defines predictive AI broadly as statistical or machine-learning models that classify or produce risk scores.
This is not evidence that every deployed system works well. It is evidence that medical AI is already part of health-system infrastructure, with clear differences by hospital size, ownership, location, and system affiliation.
[Sources]
Not an LLM-only field
Healthcare AI is built from many model families working on many kinds of signals.
IMAGES classification · detection · segmentation · reconstruction
WAVEFORMS time-series models · anomaly detection · forecasting
RECORDS risk models · survival analysis · causal estimation
LANGUAGE information extraction · retrieval · generation
MOLECULES graph models · structure prediction · generative design
ACTIONS optimization · control · robotics · reinforcement learning
Connect this slide back to Week 2. Computer vision dominates many authorized medical devices, while language models are moving into documentation and information access. Time-series models work on physiologic signals, graph models on molecules, and control methods on robots and treatment policies.
Students should resist collapsing all healthcare AI into chatbots. The model family should follow the data structure, clinical action, and acceptable error pattern.
[Sources]
A better way to scan the field
Every healthcare setting contains a different combination of sensing, judgment, and action.
SETTING AI CAN HELP HUMAN EXPERTISE
HOME monitor, coach, detect change context, goals, escalation
CLINIC summarize, screen, recommend examine, interpret, decide
HOSPITAL forecast, prioritize, coordinate respond, negotiate, recover
LAB + IMAGING find, quantify, compare validate, diagnose, communicate
RESEARCH search, simulate, generate form hypotheses, test, explain
Use this table to show that the same technical capability changes meaning by setting. An anomaly detector at home must decide when to escalate. In a laboratory, it may prioritize slides. In a hospital, it may trigger a rapid-response workflow.
The human column is not ceremonial oversight. It names expertise that the system depends on to establish meaning, act safely, and recover when the world does not match the model.
[Sources]
Six pathways to value
AI creates healthcare value in more ways than “better diagnosis.”
1 EARLIER detect risk before crisis
2 MORE PRECISE tailor care to the person
3 MORE CAPACITY return time to patients
4 MORE CONTINUITY connect visits and settings
5 MORE REACH move expertise to the point of need
6 FASTER SCIENCE search biological possibility
This is the lecture roadmap. Each pathway has mature examples, emerging systems, and speculative frontiers. The goal is to help students generate a broad opportunity inventory without confusing possibility with proof.
Ask which pathway changes the patient experience, the clinician experience, or the economics of care.
[Sources]
01
See sooner AI can extend perception across images, signals, records, and time.
This section covers perception and early detection. Emphasize that earlier is valuable only when an actionable pathway exists after the signal.
[Sources]
Medical imaging
AI can turn every scan into a second search, a measurement tool, and a comparison with the past.
FIND Small lesions, fractures, bleeds, and vascular occlusions.
MEASURE Volumes, boundaries, growth, perfusion, and treatment response.
PRIORITIZE Move time-sensitive studies to the front of a worklist.
RECONSTRUCT Improve images from faster or lower-dose acquisition.
The opportunity is not one “AI radiologist.” It is a collection of specialized perceptual tools across the imaging workflow.
The FDA’s list of AI-enabled devices shows how central imaging has been to medical AI. Explain the difference between classification, localization, segmentation, quantification, triage, and image reconstruction.
A tool that detects a possible stroke and reprioritizes the worklist creates value through time, not only diagnostic accuracy.
[Sources]
Evidence · Breast-cancer screening
The MASAI trial moved AI mammography from benchmark accuracy toward population outcomes.
TRIAL 100,000+ Women randomized to AI-supported or standard double reading in Sweden.
WORKLOAD 44.3% Lower screen-reading workload in the interim safety analysis.
OUTCOME 12% Lower interval-cancer rate reported in the 2026 follow-up.
Why it matters The system changed how radiologists allocated attention and was evaluated against what happened after screening, not only against a labeled image set.
The 2023 interim analysis reported a 44.3 percent reduction in screen-reading workload. The 2026 primary follow-up reported a 12 percent lower interval-cancer rate with AI-supported screening.
This is a stronger evidence pattern than a retrospective benchmark because it tests a clinical workflow and a patient-relevant downstream outcome. Results still come from a specific screening program, population, system, and implementation.
[Sources]
Evidence · Specialty access in primary care
Autonomous eye screening changed who completed care, not only who received an accurate score.
USUAL REFERRAL 22% completed a diabetic eye exam within six months
POINT-OF-CARE AI 100% completed the eye exam during the study visit
Among participants with an abnormal AI result, 64% completed specialist follow-up. The trial enrolled a diverse group of youth with diabetes.
The ACCESS trial is useful because it measured care-gap closure. All 81 participants in the intervention arm completed a point-of-care eye exam, compared with 18 of 82 in the control arm within six months. Among 25 abnormal AI results, 16 completed specialist follow-up.
The earlier pivotal trial established diagnostic performance in adults. ACCESS asks a different question: can autonomous AI change access and follow-through?
[Sources]
Computational pathology
A digital tissue slide can become searchable evidence, a second opinion, and a source of new biomarkers.
2.3M slides learn visual representations
+ 700K reports connect tissue to diagnostic language
→ PRISM2 cancer detection · report completion · biomarker and survival representations
One general model can support many downstream tasks, but every intended clinical use still needs its own validation.
PRISM2 was trained with 2.3 million whole-slide images and clinical question-answer supervision derived from roughly 700,000 reports. The paper reports prompt-based performance matching or exceeding several clinical-grade products on selected cancer-detection tests.
The opportunity is broader than a single classifier. The same representation may help with triage, report completion, cancer subtyping, biomarker discovery, survival research, and second opinions. The study does not establish that all of these uses are ready for routine care.
[Sources]
Signals outside the hospital
Continuous sensing can move medicine from occasional snapshots toward early change detection.
SECONDS ECG · oxygen · glucose detect rhythm, hypoxia, or metabolic change
HOURS sleep · movement · respiration identify a changing baseline
DAYS symptoms · medication · behavior forecast deterioration or recovery
The technical challenge is finding the signal. The care challenge is deciding when, where, and to whom the signal should escalate.
Sensor-based devices can support continuous measurement and longitudinal baselines. AI can detect anomalies, estimate trends, and compress large streams into actionable summaries.
False alerts, missingness, device adherence, baseline differences, and escalation capacity determine whether continuous sensing creates reassurance or noise.
[Sources]
Precision medicine
AI can connect the patient in front of us to patterns spread across millions of measurements.
PERSON goals · symptoms · history · environment
PHENOTYPE images · labs · physiology · response
MOLECULAR genome · transcriptome · proteome · microbiome
POPULATION similar trajectories · rare cases · treatment outcomes
The opportunity is not merely predicting risk. It is choosing the right intervention, dose, sequence, and follow-up for this patient.
Explain why healthcare is moving toward multimodal models. No single modality fully describes a patient. Combining clinical history, imaging, physiology, molecular data, and treatment response can support more specific hypotheses.
Distinguish association from treatment effect. A model that predicts who will do poorly does not automatically identify which treatment will help.
[Sources]
02
Act sooner The highest-value prediction is often the one that creates useful lead time.
This section moves from perception to action. The recurring question is whether the system creates enough lead time for a person or care team to change the outcome.
[Sources]
Clinical prediction
Prediction matters when it moves a decision into a window where action can still help.
TOO EARLY weak signal, many false alarms
ACTION WINDOW detect → verify → intervene
TOO LATE high certainty, little time to change course
sepsis clinical deterioration stroke readmission medication harm
Risk models trade earlier warning against uncertainty. The useful horizon depends on what the care team can do, how fast it can respond, and the harms of unnecessary intervention.
Have students name the action associated with each example. If no action exists, the prediction may create anxiety or alert burden rather than value.
[Sources]
A 2026 milestone · Sepsis
An AI early-warning system received FDA clearance after years of work on the model and the bedside workflow.
STREAM EHR data are monitored before clinicians suspect sepsis.
→ ALERT The system creates lead time and presents relevant evidence.
→ TEAM Clinicians evaluate, confirm, and begin treatment.
→ OUTCOME Johns Hopkins reports lower mortality across deployed hospitals.
The breakthrough was not a score alone. It was a monitored path from signal to response.
Johns Hopkins reported FDA clearance for its sepsis early-warning system in May 2026. The university reports that the system can identify risk before clinician suspicion and that deployed use has been associated with earlier recognition and lower mortality.
Treat institutional outcome claims carefully. The teaching point is the translation pathway: longitudinal data, prospective workflow integration, clinician response, monitoring, and regulatory review.
[Sources]
Decision support
A risk score is not a treatment plan.
MODEL OUTPUT 0.72 risk of deterioration
≠ CLINICAL ACTION repeat labs · examine · transfer · treat · watch
What evidence drove the score?
Which action is appropriate for this patient?
Who has the authority and capacity to respond?
What happens when the model is wrong?
A prediction estimates an outcome under the patterns in the data. It does not automatically estimate the effect of each possible intervention. Treatment selection may require causal evidence, clinical guidelines, contraindications, patient preferences, and resource knowledge.
This distinction prevents a common category error: treating a high-risk label as if it directly specifies the next best action.
[Sources]
Thresholds become workload
The same model can feel brilliant or unbearable depending on where the alert threshold is set.
HIGHER THRESHOLD fewer alerts more missed cases higher urgency per alert
LOWER THRESHOLD more alerts more cases detected greater review burden
Choose the threshold with prevalence, staffing, intervention cost, and patient harm in view.
This revisits Week 4 without returning to heavy mathematics. Lowering a threshold usually catches more true cases and also creates more false positives. In a low-prevalence setting, even a strong model can produce many alerts that require review.
Threshold design is therefore an organizational decision about attention and intervention, not merely a model setting.
[Sources]
Clinical copilots
A useful clinical copilot can gather evidence before it tries to generate an answer.
RETRIEVE history, trends, medications, guidelines
→ ORGANIZE timeline, problem list, contradictions
→ PROPOSE questions, differential, next steps
→ VERIFY clinician checks evidence and fit
Best near-term role: compress search and synthesis while keeping evidence visible.
Harder role: autonomous diagnosis or treatment across open-ended cases.
Large multimodal models can summarize records, answer questions, draft differentials, and retrieve guidance. The WHO guidance emphasizes governance, transparency, validation, and accountability for health uses.
Separate tasks with inspectable evidence from open-ended autonomous decision making. A copilot that reduces search time can be valuable without being authorized to make the final decision.
[Sources]
Nursing and bedside care
AI can help nurses see the whole ward without losing sight of the person in the room.
ACROSS THE WARD identify deterioration
prioritize rounds
forecast staffing and supplies
coordinate handoffs
ATTENTIONis the scarce resource
AT THE BEDSIDE interpret context
notice subtle change
explain and reassure
advocate and escalate
Nursing combines continuous surveillance, physical care, coordination, education, and advocacy. AI can help aggregate signals and prioritize attention, but it must not convert bedside work into blind compliance with a queue.
The broader opportunity is team cognition: helping the whole unit maintain a shared, current picture of patient needs.
[Sources]
Frontier · Surgical autonomy
Robotic surgery is moving from precise teleoperation toward systems that can interpret and execute parts of a procedure.
0 MANUAL human controls tools
1 ASSISTED stabilize and constrain
2 TASK automate a defined step
3 PROCEDURAL adapt across a sequence
4 FUTURE broader autonomous operation
In 2025, SRT-H completed a lengthy phase of gallbladder removal on realistic ex vivo tissue, responding to voice guidance and adapting during the procedure.
SRT-H used hierarchical policy learning and surgical video demonstrations to perform a multi-step phase of cholecystectomy on lifelike ex vivo tissue. It is an important autonomy milestone, but it did not demonstrate routine autonomous surgery on human patients.
The long-term promise includes consistency, precision, training, and access. The research path must address live-tissue variability, bleeding, rare complications, responsibility, and graceful handoff.
[Sources]
Rehabilitation and assistive care
AI can make therapy responsive to the person’s movement, fatigue, and progress.
SENSE motion · force · speech · effort
→ INTERPRET quality · compensation · fatigue
→ ADAPT difficulty · feedback · assistance
→ LEARN longitudinal progress
prosthetic control gait support speech therapy stroke recovery personalized exercise
Rehabilitation is a natural closed-loop setting. The system senses performance, estimates state, changes assistance or difficulty, and observes the response.
The meaningful outcome is not whether the model labels a movement correctly. It is whether the person gains function, confidence, and independence without losing appropriate challenge or agency.
[Sources]
03
Give time back Some of healthcare’s biggest gains may come from removing friction around care.
This section turns to the economics of attention. Many clinicians do not need AI to replace their judgment. They need systems that stop consuming expert time with repetitive documentation, routing, and coordination.
[Sources]
Evidence · Ambient documentation
AI scribes show how automation can restore the human side of a clinical encounter.
51.9% clinicians reporting burnout before use
→ 38.8% after 30 days of ambient AI use
Participants also reported less note-related cognitive load, less after-hours documentation, and more focused attention on patients.
This was a voluntary pre/post quality-improvement study, not a randomized estimate of long-term causal effects.
Olson and colleagues studied 263 physicians and advanced-practice practitioners with direct patient care. Reported burnout fell from 51.9 to 38.8 percent after 30 days. Several other experience measures improved.
The design cannot prove that the tool caused every change. Randomized trials published in 2025 also reported modest positive physician outcomes, with variation by product and endpoint.
[Sources]
Administrative and communication work
The highest-volume healthcare AI may live in work patients rarely see.
BEFORE schedule · verify coverage · prepare chart · predict no-show
DURING transcribe · retrieve · code · place orders · explain
AFTER summarize · route inbox · reconcile · follow up · bill
Automating friction can increase capacity without pretending that relationships, judgment, and responsibility are clerical tasks.
Federal hospital survey data show rapid growth in predictive AI for billing and scheduling. Generative systems are also entering inbox management, chart preparation, coding support, prior-authorization drafting, and patient communication.
These applications deserve clinical scrutiny because operational decisions affect access, delay, workload, and who receives attention.
[Sources]
Hospital operations
A hospital is a dynamic system of beds, people, equipment, queues, and uncertain arrivals.
ED arrivals discharges staffing OR cases bed status transport
FORECAST
open capacity sequence cases assign teams reduce boarding route supplies plan discharge
Prediction estimates arrivals, discharges, length of stay, demand, and staffing pressure. Optimization translates forecasts into schedules, assignments, and routing decisions under constraints.
This is a setting where classic operations research, simulation, and machine learning may matter more than generative AI. The objective must balance throughput with safety, fairness, continuity, and staff burden.
[Sources]
Medication and pharmacy
AI can follow a medication across prescribing, dispensing, adherence, and response.
PRESCRIBE interaction and dose support
→ DISPENSE verification and inventory
→ TAKE reminders and adherence patterns
→ MONITOR side effects and effectiveness
polypharmacy review antimicrobial stewardship pharmacovigilance personalized dosing
Medication safety depends on context distributed across the record, pharmacy, laboratory values, patient behavior, and symptoms. AI can support interaction checks, reconciliation, dosing, adherence, and post-market signal detection.
The human factors problem is prioritization. A system that flags every theoretical interaction can bury the alert that matters.
[Sources]
Care at home
AI can make the home a place of recovery, observation, and early support rather than a blind gap between visits.
OBSERVE symptoms, movement, sleep, vitals, medication
COACH rehabilitation, self-management, education
CONNECT summaries and escalation to the care team
ADAPT support based on changing goals and capacity
The future is not constant surveillance. It is meaningful sensing with clear boundaries and a reliable path to help.
Home-based AI can support chronic disease, post-operative recovery, rehabilitation, aging, and hospital-at-home programs. It can compress continuous data into trends and exceptions that a care team can review.
Connect to Week 5: collection should be proportional to the care purpose, and patients need to know what is sensed, who responds, and what happens if the system fails.
[Sources]
04
Expand reach AI can move parts of expertise across distance, language, and institutional boundaries.
This section asks how AI might expand access without exporting poor-quality care or imposing one institution's assumptions on every community.
[Sources]
Distributed expertise
AI can carry a specialized capability to the point where a patient first enters the system.
SCREEN LOCALLY Autonomous eye screening brings a bounded specialist task into primary care.
GUIDE ACQUISITION AI can coach a local user to capture a usable image, scan, or signal.
TRIAGE IMMEDIATELY Urgent findings can reach the right specialist before the normal queue.
CONSULT REMOTELY Decision support can help a local team collaborate with scarce expertise.
LOCAL ENTRY → AI performs or supports a bounded capability → SPECIALIST WHEN NEEDED
Use the eye-screening example to show the structural opportunity. AI does not have to replace an entire specialty. It can perform or support a bounded capability locally, help acquire usable data, prioritize urgent cases, and reserve specialist time for confirmed or complex cases.
Access still depends on equipment, connectivity, referral capacity, payment, follow-up, and trust. The goal is to move capability closer without lowering clinical standards.
[Sources]
Communication and comprehension
Language models could make medical information easier to find, translate, and understand.
PATIENT plain-language instructions · preparation · follow-up
CLINICIAN record search · evidence retrieval · handoff summaries
TEAM translation · role-specific views · shared plans
COMMUNITY public-health communication · local-language materials
Good communication is not simpler text alone. It is accurate, culturally appropriate, actionable, and open to questions.
Large language models can transform one source into different reading levels, languages, and role-specific summaries. This could reduce information barriers and help patients prepare questions.
The WHO cautions that health language models can produce false, biased, incomplete, or culturally inappropriate content. Local review, source grounding, and accessible escalation remain essential.
[Sources]
Population health
Public-health AI can find patterns that no single clinic can see.
LOCAL SIGNALS laboratory results · symptoms · wastewater · mobility · news · genomics
→ POPULATION MODEL detect · forecast · map · prioritize
→ PUBLIC ACTION investigate · communicate · allocate · intervene
CDC reported 103 internal AI use cases as of the end of 2025, spanning outbreak prevention and operational efficiency.
CDC describes AI applications for outbreak detection, data processing, forecasting, operational efficiency, and public communication. Its 2026 vision page reports 103 agency AI use cases as of December 31, 2025.
Open-source intelligence systems can also identify outbreaks of unknown cause earlier than traditional reports. Population-level systems raise distinct questions about representativeness, surveillance, false signals, and public legitimacy.
[Sources]
05
Accelerate science AI can search biological spaces that are far too large for human trial and error alone.
This section widens the time horizon. Clinical AI changes today's decisions. Scientific AI can change which treatments exist years from now.
[Sources]
AI for biology
AlphaFold changed protein structure from a scarce experimental result into a widely available prediction.
200M+
open protein-structure predictions Researchers can begin with a plausible 3D structure, form hypotheses, compare proteins, and target experiments more efficiently.
A prediction is not a complete account of biological function, interaction, or clinical effect.
The AlphaFold Database provides open access to more than 200 million predicted protein structures. EMBL reported in 2025 that it had been used by more than three million people in over 190 countries.
The achievement illustrates a different form of healthcare AI: compressing a difficult scientific inference so that experiments can start from a better hypothesis.
[Sources]
Drug and therapy discovery
AI can redesign the discovery loop from target identification through clinical development.
UNDERSTAND disease mechanism and target
→ GENERATE molecules, proteins, and candidates
→ PREDICT binding, toxicity, efficacy, manufacturability
→ TEST simulation, laboratory, animal, clinical
↺ AI can make each cycle faster. Biology still decides what survives the next test.
AI can support target discovery, structure prediction, candidate generation, property prediction, trial matching, and safety surveillance. The central advantage is search: ranking or generating promising candidates before expensive experiments.
Avoid claims that AI eliminates experimentation. In biology, models are most powerful when tightly coupled to iterative laboratory and clinical tests.
[Sources]
Frontier · Virtual patients
Digital twins could let care teams explore possible futures before choosing an intervention.
PATIENTcurrent state
treatment A → projected response
treatment B → projected response
wait + monitor → projected response
Frontier, not settled practice: credible digital twins require causal models, dense longitudinal data, uncertainty estimates, and prospective validation.
Digital-twin language covers many levels of ambition, from simple patient-specific simulations to dynamic causal models. The compelling idea is to evaluate alternative trajectories before acting.
The hard problem is counterfactual validity. Observing that people like this patient had an outcome does not prove what this patient would experience under each treatment.
[Sources]
What changed in 2025 and 2026
Several frontiers moved from impressive demos toward larger models, real workflows, and stronger evidence.
2025 Ambient AI randomized trials documentation tools tested in live clinical practice
2025 SRT-H surgical autonomy multi-step procedure on realistic tissue
2026 MASAI follow-up population screening outcome beyond image accuracy
2026 PRISM2 + sepsis clearance larger multimodal pathology and a translated early-warning system
This slide is a current snapshot. It highlights a pattern rather than claiming the field is solved: prospective trials for ambient documentation, more capable surgical research systems, longer-term randomized screening outcomes, larger pathology foundation models, and regulatory translation for a sepsis warning system.
Each milestone has a different evidence level. Students should learn to distinguish a research demonstration, a clinical trial, a regulatory authorization, and a routine-care outcome.
[Sources]
The evidence base is changing
Medical AI trials are increasingly favorable, but the field still needs broader and more rigorous evaluation.
113 randomized trials identified
82.3% reported favorable primary outcomes
67.3% were single-center
71.7% did not report adherence to a reporting guideline
Promising results are accumulating. Generalizability, trial quality, and geographic concentration remain open problems.
Wang and colleagues analyzed 4,667 primary medical AI studies and identified 113 randomized trials. Most RCTs reported favorable outcomes, but most were single-center and many did not report adherence to a guideline.
This is a productive middle position: clinical evidence is no longer absent, and it is not yet deep or distributed enough to support broad claims about every setting.
[Sources]
The translation gap
A strong model is the beginning of a healthcare system, not the end.
MODEL inputs → prediction
+ INTERFACE evidence → understanding
+ WORKFLOW alert → action
+ ORGANIZATION roles → authority
+ LEARNING monitor → improve
Clinical value appears only when the whole chain works under real conditions.
Connect directly to Week 1. An accurate model can fail because data arrive late, alerts go to the wrong person, staff lack capacity, the interface hides evidence, or the organization cannot change course.
DECIDE-AI focuses on early live clinical evaluation because human factors, workflow, and system use cannot be learned from retrospective model testing alone.
[Sources]
A practical evaluation ladder
Healthcare AI should earn trust through progressively harder evidence.
1 TECHNICAL Does it work on held-out data?
2 EXTERNAL Does it travel across sites and groups?
3 SILENT Does it behave on live data without affecting care?
4 WORKFLOW Can people use it safely and consistently?
5 OUTCOMES Does it improve care, capacity, equity, or cost?
6 LIFECYCLE Does it remain useful as the world changes?
This ladder is a teaching synthesis of clinical evaluation and lifecycle guidance. It moves from retrospective technical performance to external testing, silent prospective evaluation, live workflow, patient and system outcomes, and post-deployment monitoring.
Not every low-risk tool requires the same trial. The strength of evidence should match the consequence and the claim being made.
[Sources]
Human-AI collaboration
The strongest care team combines machine breadth with human depth.
AI CONTRIBUTES continuous monitoring
large-scale comparison
consistent measurement
rapid retrieval and simulation
+
PEOPLE CONTRIBUTE physical examination
context and causal judgment
goals, values, and consent
responsibility and recovery
Together: earlier attention, better questions, more personalized action, and more time for care.
This is not a generic human-in-the-loop claim. Name the complementary capabilities. Machines can scan large streams and compare cases. Clinicians, patients, families, and teams contribute embodied evidence, local context, goals, tradeoffs, and accountable action.
Evaluate the joint system, including what happens when people disagree with the model, the model misses a case, or the clinical environment changes.
[Sources]
Opportunity selection
Start where AI can remove a real constraint and the care system can act on the result.
VALUE IF SOLVED
HIGH VALUE · LOW READINESS RESEARCH build evidence and infrastructure
HIGH VALUE · HIGH READINESS BUILD HERE clear action, usable data, capable owner
LOW VALUE · LOW READINESS AVOID technology searching for a problem
LOW VALUE · HIGH READINESS QUESTION easy automation, weak patient benefit
ABILITY TO DEPLOY SAFELY
Use this as a short in-class inquiry. Ask students to place one healthcare AI idea on the matrix and defend both dimensions.
High readiness includes data availability, a clear user, an action pathway, integration capacity, a measurable outcome, and manageable risk. High value includes patient benefit, capacity, equity, or meaningful cost reduction.
[Sources]
The speculative horizon
Some of the most extraordinary healthcare applications may still be ahead.
PERSONALIZED MEDICATION Drugs and doses designed for one body.
ADAPTIVE TREATMENT Therapy adjusts continuously to response.
PRE-SYMPTOM DETECTION Disease found years before symptoms.
PATIENT DIGITAL TWINS Treatments tested in simulation first.
EXPERT CARE ANYWHERE Specialist guidance in any clinic or home.
AUTONOMOUS PROCEDURES Robots handle bounded tasks under supervision.
REGENERATIVE DESIGN AI designs proteins, cells, tissues, and organs.
NEURAL PROSTHETICS Devices learn intention and restore function.
SELF-RUNNING LABS AI plans, runs, and learns from experiments.
GLOBAL OUTBREAK RADAR Threats found before crises spread.
These are research directions, not promises. Their value will depend on evidence, access, governance, and the care systems built around them.
This slide is intentionally broad and future-facing. It gives students a concise inventory of major directions being theorized or actively researched, from personalized therapeutics and digital twins to autonomous procedures and self-running laboratories.
Present these as possibilities rather than forecasts. Some build directly on current advances in sensing, protein modeling, robotics, multimodal models, and public-health surveillance. Others require major scientific, clinical, economic, and governance breakthroughs.
[Sources]
A healthcare AI opportunity canvas
For any setting, ask six questions. 1 What human need or system constraint are we changing?
2 What signal, pattern, or search can AI handle unusually well?
3 What new action becomes possible, and who takes it?
4 Which expertise, relationship, and responsibility must remain human?
5 What evidence would show real clinical or system value?
6 How will the system learn safely after deployment?
SEESOONER FURTHER BETTER
Close on possibility with discipline. Healthcare offers thousands of applications because every stage contains information, coordination, uncertainty, and constrained expertise.
The durable takeaway is that AI can help medicine see sooner, reach farther, personalize more deeply, return time to care, and accelerate science. The work is to connect each capability to a real need, an action, an evidence plan, and a responsible team.
[Sources]