Precision medicine and AI in healthcare: biomarker discovery, clinical decision support

Intuition Beginner

Precision medicine matches the right treatment to the right patient at the right time, replacing one-size-fits-all care. The NIH All of Us Research Program (the Precision Medicine Initiative, 2015) is enrolling one million Americans to map individual variation in genes, environment, and lifestyle. AI accelerates the shift — algorithms now flag breast cancer on mammograms more accurately than radiologists, predict clinical deterioration hours before it strikes, and surface drug targets inside genomic data. The promise is tempered by risk: Ziad Obermeyer showed a widely used algorithm under-referred Black patients because it mistook health care spending for illness, and Black patients spend less owing to access barriers. Eric Topol argues AI should give doctors time for patients, not replace them.

Visual Beginner

Domain	AI task	Landmark result	See
Radiology	Mammography	Better than radiologists (McKinney 2020)	33.07.*
Pathology	Cancer grading	Computational pathology	35.03.03
Dermatology	Skin lesion classification	CNN matches dermatologists (Esteva 2017)	29.03.*
Ophthalmology	Diabetic retinopathy	IDx-DR, first autonomous AI diagnostic (FDA 2018)	35.03.04
Cardiology	Arrhythmia on ECG	Apple Watch AFib detection	35.03.02
EHR	Sepsis, readmission, deterioration	Deep learning on EHRs (Rajkomar 2018)	35.01.02

Precision medicine application	Example	See
Companion diagnostic	HER2 test before trastuzumab	35.03.03
Pharmacogenomics	Warfarin dosing by genotype	35.07.02
Precision oncology	Tumor-board molecular profiling	35.08.02
Risk stratification	Polygenic risk scores	35.08.02

Worked example Beginner

Worked example: a clinical decision support alert for sepsis

A 68-year-old man is on the surgical ward two days after bowel surgery. An EHR-integrated model recomputes his sepsis risk every fifteen minutes from vital signs, labs, and nursing notes (see 35.02.02, bacterial pathogenesis).

Step 1. At 03:00 the model lifts his risk from 4% to 27% as heart rate climbs and lactate ticks upward — no single value crosses a threshold, but the joint trajectory does.

Step 2. The nurse draws blood cultures and the resident starts broad-spectrum antibiotics within the hour, hours before shock would trigger standard early-warning scores.

Step 3. The alert is probabilistic, not diagnostic. Many are false positives, feeding alert fatigue; safe use needs external validation, calibration to local case mix, and subgroup bias monitoring (see 35.06.02).

Check your understanding Beginner

Question 1: Precision medicine chiefly aims to:

A) Use the newest available drugs B) Tailor prevention and treatment to individual variability C) Replace physicians with algorithms D) Sequence every patient's genome regardless of need

Answer: B. Collins and Varmus frame precision medicine around individual variability in genes, environment, and lifestyle ^{[NEJM 372 (2015) 793-795]}.

Question 2: The Obermeyer algorithm under-referred Black patients because it:

A) Used race as an explicit input B) Used health care cost as a proxy for need C) Was trained only on White patients D) Contained a software bug

Answer: B. Cost is a biased proxy: Black patients spend less at equal illness owing to access barriers ^{[Science 366 (2019) 447-453]}.

Question 3: True or false: a model with high AUC and good calibration can still give biased clinical decisions.

Answer: True. Discrimination and calibration ignore subgroup fairness; threshold and proxy choices can encode inequity even in an accurate model.

Formal definition Intermediate+

Precision medicine. Collins and Varmus define it as an approach to treatment and prevention that accounts for individual variability in genes, environment, and lifestyle ^{[NEJM 372 (2015) 793-795]}. The All of Us cohort of one million participants supplies the variation data (see 35.08.02, genomic medicine).

Biomarkers. A biomarker is an objectively measured indicator of normal or pathogenic biology or of therapeutic response — HER2 for trastuzumab eligibility (see 35.03.03, cancer), HbA1c for glycemic control (see 35.03.04, metabolic syndrome), troponin for cardiac injury (see 35.03.02, cardiovascular disease). Performance rests on sensitivity (TP/(TP+FN)), specificity (TN/(TN+FP)), and AUC — the area under the receiver operating characteristic curve, equal to the probability that a random case outranks a random non-case, ranging from 0.5 (chance) to 1.0 (perfect). Positive predictive value tracks prevalence:

$$\text{PPV}=\frac{\text{sens}\cdot \pi }{\text{sens}\cdot \pi +(1-\text{spec})\cdot (1-\pi )}$$

so a strong biomarker in a common disease floods the clinic with false positives when prevalence is low (see 29.01.03, statistical reasoning).

Clinical decision support. A clinical decision support (CDS) system delivers patient-specific assessments — sepsis, deterioration, or readmission risk learned from electronic health records (see 35.01.02, homeostasis). Deep models now train on hundreds of thousands of records ^{[NPJ Digital Medicine 1 (2018) 18]}, but deployment demands calibration: predicted probabilities must match observed event rates. Discrimination and calibration are independent — a model can rank correctly yet be numerically wrong, so both must be re-checked on external populations under distribution shift (see 35.06.02).

Key result Intermediate+

The cost-as-proxy bias in population-health algorithms

The nontrivial question is how an accurate model can still discriminate. A widely deployed algorithm ranked patients for high-risk care management using predicted annual health care cost $C$ as its proxy for health need $I$ ^{[Science 366 (2019) 447-453]}. Obermeyer et al. found that, at any fixed cost level, Black patients were substantially sicker than White patients.

Setup. Suppose cost factors as need times access, $C=I\cdot A$, where $A$ is the rate at which a patient converts health need into health care spending.

Claim. Flagging patients by a cost threshold $C^{\ast }$ under-refers Black patients whenever their access $A$ is lower.

Proof. A patient is flagged when $C\ge C^{\ast }$, equivalently $I\cdot A\ge C^{\ast }$, i.e. $I\ge C^{\ast }/A$. Because Black patients face access barriers, $A_{\text{Black}}<A_{\text{White}}$, so the illness required to clear the threshold satisfies $C^{\ast }/A_{\text{Black}}>C^{\ast }/A_{\text{White}}$. At the threshold a Black patient must therefore carry a higher illness burden than a White patient to be flagged, and equally ill Black patients go unflagged — exactly the gap Obermeyer measured ^{[Science 366 (2019) 447-453]}. Replacing cost with a direct illness proxy (active chronic conditions) cut the gap nearly in half.

The lesson is that the choice of proxy, not the model architecture, encoded racial inequity.

Exercises Intermediate+

Exercise 1 (easy). Define precision medicine in your own words and give one example each of a genomic, an environmental, and a lifestyle input an oncologist might use.

Exercise 2 (medium). A biomarker has sensitivity 0.95 and specificity 0.90 for a disease with prevalence 1%. Using the PPV formula, compute the positive predictive value and explain why so few flagged patients are true positives.

Exercise 3 (medium). Restate the Obermeyer bias in the $C=I\cdot A$ framing. Why does replacing cost with an active-chronic-condition count reduce the racial gap?

Exercise 4 (hard). A sepsis-prediction model achieves AUC 0.85 internally but 0.62 at another hospital. List four reasons for the drop and propose a prospective validation protocol before clinical release.

Exercise 5 (hard). Contrast discrimination, calibration, and algorithmic fairness. Construct a case in which a model satisfies all three yet is still unsafe to deploy, and state what additional safeguard you would require.

Advanced results Master

P4 medicine and digital twins

Leroy Hood's P4 medicine — predictive, preventive, personalized, participatory — folds multi-omics into systems models, moving beyond single-biomarker thinking (see 20.05.04, reductionism in biology). A digital twin is a personalized physiological model that simulates disease progression and drug response for one individual (see 35.01.02, homeostasis; 43.*, numerical methods). Integration of genomic, transcriptomic, proteomic, and metabolomic data is the substrate these twins run on (see 35.08.02).

Explainability, liability, and privacy

Clinicians need explainable AI (XAI) to trust a recommendation; the black-box problem clashes with informed consent and medical malpractice liability when a model errs (see 20.02., ethics; 20.08.02, scientific realism). Health data is governed by HIPAA and GDPR, but wearable and app streams leak outside those regimes into surveillance-capitalism business models (see 36., media literacy; 30.02.03, Zuboff). Genetic prediction raises further consent and privacy questions under GINA (see 20.02.06, AI ethics).

AI for global health and the future

Smartphone diagnostics and AI chest X-ray for tuberculosis serve low-resource settings where specialists are scarce (see 31.06., anthropology, global health; 35.06., public health). IDx-DR (2018) was the first fully autonomous AI diagnostic. Ahead: AI mental-health chatbots such as Woebot and Wysa (see 29.10.02, therapy), senolytics and aging biomarkers (see 31.04., biological anthropology), brain-computer interfaces (see 20.06., consciousness; 29.02.*, neuroscience), and quantum computing for drug discovery and protein folding (see 20.03.02, philosophy of quantum mechanics). The arc points toward safety-critical AI that must remain under human governance (see 20.02.06, Bostrom).

Connections Master

Genomic medicine

Biomarkers and risk scores originate in genomic data — GWAS variants, tumor sequencing, pharmacogenes (35.08.02, genomic medicine).

Pharmacology and oncology

Companion diagnostics gate targeted therapy (HER2/trastuzumab; 35.03.03, cancer), while pharmacogenomics personalizes dosing for drugs such as warfarin (35.07.02, pharmacokinetics; 35.07.03, drug classes).

Chronic disease

Cardiovascular risk models (35.03.02), metabolic closed-loop systems such as the artificial pancreas (35.03.04), and cancer molecular profiling (35.03.03) all feed clinical decision support.

Health disparities, ethics, and data

Proxy bias, Eurocentric training data, and surveillance of wearable streams place AI medicine inside the political economy of health (35.06.02, health disparities; 30.04.03, race and ethnicity; 20.02.06, AI ethics; 36.*, media literacy; 30.02.03, Zuboff).

Computing, perception, and modeling

Deep-learning foundations (33.07., computing), image-recognition analogies in visual art (29.03., perception; 34.03., visual art), and numerical modeling (43.) underpin these systems.

Upstream link

Genomic medicine (35.08.02) supplies the variant catalogs and risk models that precision medicine deploys, and the future-of-medicine overview (35.08.01) frames this unit within genomics, AI, and global health.

Historical and philosophical context Master

The Precision Medicine Initiative (2015)

Francis Collins and Harold Varmus launched the Precision Medicine Initiative in the New England Journal of Medicine, reframing the post-genome era around individual variation rather than the average patient ^{[NEJM 372 (2015) 793-795]}. The All of Us Research Program grew from that call toward a population-scale evidence base of one million participants.

The arc of medical AI

Medical AI traveled from MYCIN in the 1970s — an expert system that never reached the bedside — through the overpromised IBM Watson era to convolutional networks that now match specialists. Esteva's skin-cancer classifier (2017), Gulshan's diabetic-retinopathy model (2016), and McKinney's mammography system (2020) marked the turnaround. In 2018 the FDA cleared IDx-DR, the first autonomous AI diagnostic — a system that can refer a patient without a physician in the loop. The trajectory bends from advisory tools toward delegated decisions, raising the stakes for validation and oversight.

Deep Medicine and the human factor

Topol's thesis in Deep Medicine is that AI's highest value is restoring time for empathy and the doctor-patient relationship: pattern recognition offloaded to machines, presence returned to clinicians ^{[Deep Medicine (Basic Books, 2019)]}. The deeper question is whether a black box can participate in a clinical encounter where understanding suffering, not only predicting it, is the point (see 20.06.*, consciousness; 31.06.02, Kleinman's illness narratives). Precision medicine's promise turns on keeping the human inside the loop, with AI as augmenter rather than replacement.

Bibliography Master

1. Collins, F. S. & Varmus, H. (2015). "A New Initiative on Precision Medicine." New England Journal of Medicine, 372(9), 793–795.

2. Topol, E. (2019). Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books.

3. Obermeyer, Z., Powers, B., Vogeli, C. & Mullainathan, S. (2019). "Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations." Science, 366(6464), 447–453.

4. Rajkomar, K., Oren, E., Chen, K. et al. (2018). "Scalable and Accurate Deep Learning with Electronic Health Records." NPJ Digital Medicine, 1, 18.

5. Gulshan, V., Peng, L., Coram, M. et al. (2016). "Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs." JAMA, 316(22), 2402–2410.

6. Esteva, A., Kuprel, B., Novoa, R. A. et al. (2017). "Dermatologist-level Classification of Skin Cancer with Deep Neural Networks." Nature, 542(7639), 115–118.

7. McKinney, S. M., Sieniek, M., Godbole, V. et al. (2020). "International Evaluation of an AI System for Breast Cancer Screening." Nature, 577(7788), 89–94.

8. Hood, L. & Flores, M. (2012). "A Personal View on Systems Medicine and the Emergence of Proactive P4 Medicine: Predictive, Preventive, Personalized and Participatory." New Biotechnology, 29(6), 613–624.

9. U.S. Food and Drug Administration. (2019). "Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD)." FDA Discussion Paper.

10. Char, D. S., Shah, N. H. & Magnus, D. (2018). "Implementing Machine Learning in Health Care — Addressing Ethical Challenges." New England Journal of Medicine, 378(11), 981–983.