Transforming Healthcare Through Intelligent Systems, Explained

UPI Study › Courses › Introduction to Artificial Intelligence › Free lessons › Lesson 9.3 Free with a free account

Transforming Healthcare Through Intelligent Systems

Introduction to Artificial Intelligence · Chapter 9: AI Applications, Ethics, and the FutureSection 9.3 · lesson 64 of 6916 min · free with a free account

← AI in Everyday Life: Current ApplicationsNext lesson: AI Across Industries: Manufacturing, Logistics, and Beyond →

Earn 3 college credits for Introduction to Artificial IntelligenceRecommended for credit by NCCRS · Computer Science 373 · self-paced · one payment, lifetime access
Enroll in Introduction to Artificial Intelligence · $250 →Download the syllabus (PDF)

In this lesson

AI in diagnosis, drug discovery and monitoring, and how to validate it.

  • Calculate sensitivity, specificity and predictive values, and explain why prevalence changes what a positive result means
  • Describe how AI supports medical imaging, drug discovery, genomics and patient monitoring
  • Explain threshold trade-offs, alert fatigue, calibration and drift in clinical models
  • Outline the stages of clinical validation, regulation and data protection for medical AI

Free study notes

Get this lesson's notes by email

The key points and 4 practice questions with answers, in your inbox. One email, free.

One email. Unsubscribe in one click.

Transcript

Everything said in the video, section by section (2,182 words). Read along, or search it.

Opening Question

What Lena's Positive Test Means

Lena takes a screening test for a disease, and it comes back positive. The test finds 90 percent of people who have the disease, and correctly clears 95 percent of healthy people. So how likely is it that Lena is actually ill? It's tempting to say ninety percent. Hold that thought.

Why Healthcare

Rare Signals in Floods of Data

Medicine is full of rare, important signals hidden in floods of data: a missed shadow on one scan out of dozens, the first signs of a patient getting worse. A new drug can take more than ten years to reach patients, and most candidates fail on the way. Finding rare signals is what machine learning does well. But healthcare is also where mistakes cost the most, so medical AI is built, tested and regulated differently. And a note: this lesson's numbers are made-up round figures, not clinical data.

Medical Imaging

Detection, Segmentation, Triage and Measurement

Imaging is where AI has made its clearest mark. Convolutional networks learn from images labeled by specialists, or by later confirmed results, like a biopsy. They read X-rays, CT and MRI scans, looking for lung nodules or bleeding in the brain. They act as an extra reader of mammograms. They scan huge pathology slides for cancer cells, check retinal photos for diabetic eye damage, and look at skin photos and heart rhythm traces. The jobs are detection, is something there? Segmentation, outlining it. Triage, moving urgent scans to the top. And quantification, measuring what people estimate by eye. In almost every case, the tool supports a clinician, who keeps responsibility for the decision.

Diagnostic Measures

Sensitivity, Specificity and Predictive Values

How do you judge a diagnostic tool? Four numbers. Sensitivity: of the people who have the disease, how many does the test find? Specificity: of the healthy people, how many does it correctly clear? Positive predictive value: of the people it flags, how many really are ill? And negative predictive value: of those it clears, how many really are healthy? Sensitivity and specificity describe the test. The predictive values describe what a result means for a patient, and they depend heavily on how common the disease is.

Worked Example

Same Test, Two Settings

Take that same test, 90 percent sensitive and 95 percent specific, and give it to ten thousand people in two settings. Screening the public, where 1 percent are ill: it finds 90 of the 100 sick people. But 5 percent of 9,900 healthy people is 495 false alarms. So only 90 of 585 positives are real: a positive predictive value of 15 percent. In a specialist clinic, where 20 percent are ill, it finds 1,800, with only 400 false alarms. Now a positive is right 82 percent of the time. That's why a screening result is a reason for a follow-up test, not a diagnosis. And a tool validated in one setting can't simply be moved to another.

Thresholds

Missed Cases Versus False Alarms

A model gives each case a score, and a threshold turns it into a yes or no. Moving the threshold trades one kind of error for the other. In the 1 percent population, a sensitive setting finds 98 of 100 cases, but raises nearly 2,000 false alarms. Its positive predictive value is under 5 percent. The balanced setting finds 90, with 495 false alarms. A specific setting finds only 73, but with just 99 false alarms, and 42 percent of its positives are real. There's no single best setting. It depends on the cost of a missed case against the cost of a false alarm: extra scans, anxiety, an unnecessary biopsy.

Quick Check

A Positive in a 10 Percent Clinic

Quick check. The same test, 90 percent sensitive and 95 percent specific, is used on 1,000 people in a clinic where 10 percent are ill. A patient there tests positive. Roughly how likely is it they're ill? Pause, and work it out. About two in three. A hundred are ill, and the test finds 90 of them. Of the 900 healthy people, 5 percent, or 45, are false alarms. So 90 of 135 positives are real: 67 percent. Nine in ten is the trap. That's the test's sensitivity, not what a positive means.

Common Pitfalls

Shortcuts, New Scanners and Automation Bias

Imaging models also go wrong in typical ways. Shortcut learning: a model picks up a clue unrelated to disease. Studies found skin-lesion models partly relying on rulers or ink marks seen beside cancers in the training photos, and chest X-ray models picking up the hospital or scanner. Different equipment or patients: a model trained on one scanner or population can do worse on another. And noisy labels: specialists sometimes disagree. Finally, automation bias. A busy clinician may accept a wrong suggestion, because the machine seemed confident.

Drug Discovery

AI at Every Stage and AlphaFold

Drug discovery is slow, costly and uncertain. It runs from choosing a biological target, to finding and improving compounds, to lab tests, human trials and approval. It often takes more than ten years, and about nine in ten candidates that enter human trials never reach approval. AI helps at every stage: mining papers to propose targets, ranking compounds, predicting toxicity, matching patients to trials, and spotting side effects after approval. One advance stands out. DeepMind's AlphaFold predicts a protein's 3D shape from its amino acid sequence, for many proteins close to lab accuracy. Finding such a shape used to take months or years of experiments.

Virtual Screening

An Enrichment of 40 Times

Virtual screening ranks compounds on a computer before any are made. Picture a million compounds, where 0.1 percent, a thousand, are truly active. Each lab test costs a dollar. Testing everything costs a million dollars. Testing a random 1 percent costs ten thousand, and finds about 10 actives. Testing the model's top 1 percent costs the same ten thousand, but finds 400. That's 4 percent active instead of 0.1: an enrichment of 40 times. But a ranking is a shortlist, not proof. Every pick must still be tested, and most that pass the lab will still fail later.

Genomics

Reading Variants and Precision Medicine

Genomics has the same flavor. A human genome is about 3 billion letters, and two people differ at only about 0.1 percent of positions. That's still millions of differences. AI reads the sequencer's output, deciding where a person's DNA differs from the reference. It predicts whether a variant is likely to disrupt a gene, or be harmless. It predicts how genes switch on and off in different tissues. And it supports precision medicine: matching treatment to a patient's genetic profile, like a tumor's genetic changes.

Polygenic Risk Scores

A Weighted Sum of Risk Copies

AI also estimates the risk of common diseases from many small variants, as a polygenic risk score. It's a weighted sum. Each variant has a weight from past studies. A person's score adds the weight once for every risk copy they carry: zero, one or two copies. Person one carries two copies of A, one of B, one of D and two of E. That adds up to 0.57. Person two carries one copy each of C and E: 0.32. So person one is estimated at higher risk. A score shifts a probability. It isn't a diagnosis. And most large genetic studies used people of European ancestry, so scores built from them tend to be less accurate for everyone else.

Patient Monitoring

Early Warnings From Health Records

Hospitals hold huge electronic health records: diagnoses, lab results, medications, notes and bedside monitor readings. Predictive models use them to see trouble coming, early enough to act. They flag patients who may soon deteriorate, or develop sepsis. They spot patients likely to be readmitted soon, and predict bed demand and missed appointments. Wearables track heart rhythm, glucose or oxygen at home, and alert a care team when readings drift. Language models draft summaries of notes, which a clinician must check. Monitoring data are time series, so filtering tracks a patient's hidden condition, and anomaly detection notices a departure from that patient's normal. A smartwatch that flags an irregular pulse is a familiar example.

Alert Fatigue

When Most Alerts Are False

But an early warning helps only if people respond. Suppose a ward model catches 20 true deteriorations a month, and each alert takes 10 minutes of nurse time. The total alerts are 20 divided by the positive predictive value. At 50 percent, that's 40 alerts a month, about 13 minutes a day. At 20 percent, 100 alerts, 33 minutes a day. At 5 percent, 400 alerts, over two hours a day, and 380 of them false. When most alerts are false, staff learn to ignore them, and the true ones get ignored too. That's alert fatigue.

Quick Check

False Alerts From a Sepsis Model

Quick check. A sepsis model on another ward catches 15 true cases a month. Only 10 percent of its alerts are real. How many false alerts does the staff handle each month? Pause, and decide. A hundred thirty-five. Total alerts are 15 divided by 0.1, which is 150, and only 15 of them are real. A hundred fifty is the tempting answer, but that's every alert, including the true ones.

Calibration and Drift

When 20 Percent Risk Means 20 Percent

A risk model should also be calibrated. When it says 20 percent risk, about 20 in 100 such patients should have the event. A model can rank patients well and still be miscalibrated. Then anyone who reads its number as a probability is misled. Models also drift. A new lab test, a coding change, a different patient mix or a new treatment can erode accuracy over time. Some widely used hospital prediction tools were checked independently, and performed well below what was first reported. So test locally, and keep monitoring.

Clinical Validation

Evidence Built in Stages

Clinical validation asks whether a tool works for the patients it will really see, and whether using it makes them better off. Evidence builds in stages. An internal test on held-back data from the same source. External validation at other hospitals, on other equipment and patients. A prospective study in live use. A randomized trial of patient outcomes, with the tool and without it. And ongoing monitoring after launch. Accuracy isn't benefit. A tool can find disease accurately and change nothing, if the disease would be found anyway, or the extra tests do more harm than good.

Subgroup Results

An Average Hides a Weak Scanner

An average can also hide a weak spot. Suppose a tool has 94 percent sensitivity on scanner A, and 78 percent on scanner B. Seventy percent of the hospital's scans come from A. Overall, that's 0.7 times 94, plus 0.3 times 78: about 89 percent. Eighty-nine sounds fine. But a patient scanned on B has a 22 percent chance of a missed finding. So report results by scanner, age, sex and ancestry.

Regulation and Data Law

FDA, EU Rules and Health Privacy Laws

Regulators treat diagnostic software as a medical device. In the United States, the FDA reviews it. In the European Union, the Medical Device Regulation applies, and the AI Act adds duties for high-risk AI. The evidence required grows with the risk to patients. And a model that keeps learning after approval needs controlled update procedures, since each change can alter its behavior. Health data reveal diagnoses, genetics and habits, and can't be changed once leaked. Laws give it special care: HIPAA in the U.S., the GDPR in Europe, and India's Digital Personal Data Protection Act.

Protecting Data

De-Identification, Federated Learning and Noise

Training a model takes a lot of such data, so protection must be designed in. De-identification removes names and ID numbers. But birth date, sex, postal code and a rare diagnosis can still point to one person. Access control and encryption limit who sees the data. Consent and minimization use only what's needed, for agreed purposes. Federated learning shares model updates, not records. Differential privacy adds noise, at some cost in accuracy. And synthetic records help testing, but can copy real patients or miss rare cases.

Earning Trust

Clinicians Stay in Charge

A tool gets used only if clinicians and patients trust it, and that trust is earned. Clinicians stay in charge: a named professional decides, and is accountable. Honest uncertainty beats a confident answer. Heat maps can help explain, but can look convincing even when the model used a shortcut. Clinicians need to keep their own skills. The tool must work fairly for every group it serves. And patients should be told when AI contributes to their care. Before trusting a tool, ask: what was it trained on? Was it tested elsewhere, by subgroup? Do patients do better with it? Who's accountable, and how will it be monitored?

Back to Lena

About 15 Percent, Not Ninety

Back to Lena. Her test is 90 percent sensitive and 95 percent specific, used for screening the public, where about 1 percent are ill. That's exactly the first setting: only about one positive in seven is real. Her chance of being ill is about 15 percent, not ninety. Her positive result is a reason for a follow-up test, and a conversation with her doctor, not a diagnosis.

Key Takeaways

Healthcare AI in Summary

To sum up. A test's predictive values depend on how common the disease is, so the same positive can mean fifteen or eighty-two percent. Thresholds trade missed cases against false alarms, and too many false alarms cause alert fatigue. In drug discovery and genomics, AI ranks and estimates, but every candidate still faces tests and trials. And validation must be external, prospective and reported by subgroup, while patient data stays protected and trust is earned with evidence.

Think About It

A 95 Percent Accurate X-Ray Tool

Here's something to think about. A vendor says its chest X-ray tool is 95 percent accurate in its own study. It flags one patient in ten, and only a few of them turn out to be ill. Is the tool failing, or is that expected? What would you ask the vendor, and what would you test locally?

About Introduction to Artificial Intelligence at UPI Study

Introduction to Artificial Intelligence (Computer Science 373) is a self-paced online course of 69 video lessons in 9 chapters, from intelligent agents and search to game strategy, neural networks, logic and reasoning, machine learning, language, robotics and AI ethics. It carries a NCCRS credit recommendation for 3 semester credits.

Every lesson transcript is free on pages like this one, and the videos are free with a free UPI Study account. Enrolling adds a practice quiz after each lesson, the graded work and final exam that earn the credit, lifetime access, and an official transcript on completion. UPI Study has helped over 130,000 students with college credits. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Enroll now → Course details Download the syllabus All Computer Science courses Every course with PRO →

Compare: AI vs Current Trends in CS and IT

Why take it here

Why students take Introduction to Artificial Intelligence with UPI Study

How machines search, reason and learnIntroduction to Artificial Intelligence covers how AI systems are built: intelligent agents, search and heuristics, game strategy, logic, neural networks, machine learning, language and perception, and the ethics of AI. It is a common upper-level course in computer science degrees.
Credit at a fraction of the priceThe same 3 credits can cost $750 or more at a US college. Here it is one payment, self-paced, with lifetime access.
Finish your degree soonerTransfer credits reduce the courses you still owe, so you can graduate earlier or lighten a heavy semester.
Strengthen an applicationA completed college-level course on an official transcript shows admissions and scholarship committees you can do the work.
Study on your own scheduleNo start dates, no deadlines. Watch the lessons, practice, and take the final when you are ready.
130,000+students helped with college creditsACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Where the credit goes

Universities that work with NCCRS

Introduction to Artificial Intelligence carries a NCCRS credit recommendation. These universities work with NCCRS; some are UPI Study's signed partners. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Thomas Edison State UniversityThomas Edison State UniversitySigned partnerUniversity of the PeopleUniversity of the PeopleSigned partnerCharter Oak State CollegeCharter Oak State CollegeSigned partnerStrayer UniversityStrayer UniversitySigned partnerColumbia Southern UniversityColumbia Southern UniversitySigned partnerIndiana TechIndiana TechSigned partnerPenn StatePenn StateWorks with NCCRSSouthern New Hampshire UniversitySouthern New Hampshire UniversityWorks with NCCRSSUNYSUNYWorks with NCCRS +many more cooperating universitiesFind my college →

Find my college → Not listed? Ask Athena in the corner of this page, or book a 15-minute call.

FAQ

Questions students ask us

About Transforming Healthcare Through Intelligent Systems

Why does a positive screening test not always mean a person has the disease?
A positive result's meaning depends on how common the disease is, not only on the test's accuracy. A test with 90 percent sensitivity and 95 percent specificity, given to 10,000 people of whom 1 percent are ill, finds 90 true cases but raises 495 false alarms, so only about 15 percent of positives are real. In a clinic where 20 percent are ill, the same test's positives are right about 82 percent of the time.
What is the difference between sensitivity and positive predictive value?
Sensitivity is the share of people with the disease whom the test finds, while positive predictive value is the share of flagged people who are truly ill. Sensitivity and specificity describe the test itself. Predictive values describe what a result means for a patient and depend heavily on how common the disease is. A frequent mistake is reading a 90 percent sensitive test as a 90 percent chance of illness after a positive result.
What is alert fatigue in clinical AI?
Alert fatigue happens when so many alerts are false that staff learn to ignore them, and true alerts get ignored too. Total alerts equal the true cases caught divided by the positive predictive value. A ward model catching 20 true deteriorations a month at a 5 percent positive predictive value produces 400 alerts, 380 of them false, which at 10 minutes each costs over two hours of nurse time a day.
How does AI help with drug discovery?
AI helps by ranking options so lab work focuses on the most promising ones. It mines papers to propose targets, ranks compounds in virtual screening, predicts toxicity, matches patients to trials and spots side effects after approval. DeepMind's AlphaFold predicts a protein's 3D shape from its amino acid sequence, for many proteins close to lab accuracy. Candidates still need lab tests and human trials, and about nine in ten that enter trials never reach approval.

About the Introduction to Artificial Intelligence course

How many college credits is Introduction to Artificial Intelligence?
Introduction to Artificial Intelligence (Computer Science 373) carries a NCCRS credit recommendation for 3 semester credits. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.
How much does Introduction to Artificial Intelligence cost and what is included?
$250, paid once, with lifetime access: all 69 video lessons, a practice quiz after each, the graded work and final exam, and an official transcript when you finish. Students who want several courses choose UPI Study PRO instead.
Which universities accept UPI Study credits?
UPI Study has signed partner agreements with universities such as Thomas Edison State University, University of the People and Charter Oak State College, and works with universities such as Penn State, SNHU and SUNY. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities. Use "Find my college" on this page to check yours.
How long does Introduction to Artificial Intelligence take?
It is self-paced with no deadlines. The 69 lessons total about 16 hours of video; most students add reading and practice time and finish in a few weeks to a few months, on their own schedule.
How are the exams done?
Quizzes and assignments are online in the course. The final exam is taken online with proctoring; a person on the UPI Study team checks assignments, usually within two days.
How do my credits reach my college?
When the course shows 100% complete, request the official transcript from your dashboard. It is sent to the college you name, and you can send it to more than one. Check your college first with "Find my college" on this page, or ask Athena.

Ready for the full course? Earn 3 credits in Introduction to Artificial Intelligence.

69 video lessons, a quiz after each one, and the graded work that earns the credit. Recommended for credit by NCCRS.

Enroll in Introduction to Artificial Intelligence →Download the syllabusEvery course with PRO →
Introduction to Artificial Intelligence
3 credits · NCCRSEnroll →