At the 939-slot boundary (K = 20%), Model B rescues 3 true cases missed by Model A, delivering a 3.0x higher marginal yield (4.92% vs 1.64%) among the 61 swapped candidates. Across the entire 939-person queue, Model B detects 55 latent cases vs 53 for Model A (+2 net cases overall).
HiddenCare compares two ways of choosing who gets screened for diabetes when a clinic cannot test everyone.
Interactive research study on public health data. Not an individual diagnosis tool or hospital-deployed system.
When screening algorithms learn only from past diagnosis records, people who rarely visit doctors can be misclassified as low risk. This project compares how training label definitions reorder screening triage priorities across a population under finite testing capacity.
How to use: Adjust the screening budget slider below to compare how both models allocate test slots and detect hidden cases.
When Health Algorithms Overlook People Who Don't Visit Doctors
A community clinic only has enough funding to test a fraction of residents (e.g., 20% of the town). Who gets invited first?
Standard AI (Model A) trains on past doctor visits. Because uninsured people rarely see doctors, their records look healthy even when they are sick.
At the 20% budget cutoff, Model B swaps 61 overlooked people into the queue, achieving a 3.0x higher hit rate (4.92% vs 1.64%, or 3 cases vs 1 case) in that marginal swapped cohort compared to Model A. Across the full 939-person queue, Model B detects 55 latent cases compared to 53 for Model A (+2 net cases overall).
Strictly previously undiagnosed adults with zero temporal overlap with training data. Biochemically confirmed 116 latent diabetes cases (HbA1c ≥ 6.5%).
Pre-pandemic baseline comparing administrative self-report labels against biomarker-augmented screening labels.
Screening Queue Triage Under Finite Capacity
Adjust screening capacity to simulate how proxy-trained versus biomarker-augmented triage prioritizes previously undiagnosed adults (N = 4,698 holdout).
Quick Budget Presets:
Cumulative yield: How many latent diabetes cases (HbA1c ≥ 6.5%) are identified as community screening capacity expands from 5% to 50%.© 2026 Cao Thien An Nguyen • GitHub: aansensei
Screening efficiency: Number of diagnostic blood tests needed to find one latent case. Lower values indicate higher testing efficiency.© 2026 Cao Thien An Nguyen • GitHub: aansensei
The 61-Person Margin: Who Swaps at the 20% Cutoff?
At exactly 20% capacity (K = 939 slots), 61 individuals differ between Model A and Model B queues. Inspect their joint rankings and individual clinical records.
Boundary swap precision: In the 61-person boundary swap at 20% budget, Model B achieves a 4.92% hit rate (3 / 61 cases) compared to Model A's 1.64% (1 / 61 cases).© 2026 Cao Thien An Nguyen • GitHub: aansensei
Joint priority ranking: Each dot represents a patient. Points above the diagonal line are prioritized higher by Model B than Model A.© 2026 Cao Thien An Nguyen • GitHub: aansensei
| SEQN | Demographics | Anthropometry | Access & Income | Priority Ranking | Observed HbA1c | Clinical Status |
|---|---|---|---|---|---|---|
| #131367 | 58y / Female | BMI 35.8 • Waist 120.5cm | Insured • Missing PIR | Rank A: #1038 • Rank B: #862 | 6.8% | Rescued Case (HbA1c ≥ 6.5%) |
| #132056 | 65y / Female | BMI 34.1 • Waist 108.8cm | Uninsured • Low PIR | Rank A: #1057 • Rank B: #915 | 8.2% | Rescued Case (HbA1c ≥ 6.5%) |
| #142034 | 45y / Male | BMI 44.2 • Waist 131.9cm | Insured • Missing PIR | Rank A: #1072 • Rank B: #829 | 6.8% | Rescued Case (HbA1c ≥ 6.5%) |
Health Equity Audit Across Insurance and Income Strata
Evaluating whether training on clinical diagnosis history systematically penalizes individuals with lower healthcare access.
1. The 3x Insurance Sensitivity Divide: At the primary 20% capacity budget, both models successfully detect ~50.0% of latent cases among insured individuals (51/102 cases for Model A, 52/102 for Model B). Conversely, screening sensitivity drops to 15.4% (Model A) and 23.1% (Model B) among uninsured individuals, leaving over 75% of uninsured cases undetected.
2. Relative Biomarker Recovery: Incorporating the laboratory biomarker (Model B) allocates 6 additional screening slots to uninsured participants and catches +1 additional true case (+50% relative increase in yield), demonstrating that biomarker augmentation partially corrects access bias.
3. Small-Sample Integrity: In the holdout cohort of previously undiagnosed adults, the uninsured subgroup holds N = 422 participants with exactly 13 latent positive cases. In accordance with biostatistical best practices, we report exact participant counts and avoid over-extrapolating small-sample estimates into sweeping algorithmic fairness claims.
Health equity audit: Compares screening sensitivity between insured and uninsured populations, highlighting where biomarker training closes access-driven gaps.© 2026 Cao Thien An Nguyen • GitHub: aansensei
Decision Curve Analysis & Net Monetary Benefit
Quantifying financial and health utility trade-offs across screening test costs and complication avoidance values.
Decision curve analysis: Evaluates net clinical and economic benefit across decision threshold probabilities compared to universal or zero screening.© 2026 Cao Thien An Nguyen • GitHub: aansensei
| Allocation Policy | Cases Found | Total Testing Cost | Cost per Case Detected | Net Monetary Benefit (NMB) |
|---|---|---|---|---|
| Random Outreach | 1011.9 | $23,475 | $1,012 | $173,725 |
| Model A (Diagnosis Only) | 53.0 | $23,475 | $443 | $427,025 |
| Model B (Combined Biomarker) | 55.0 | $23,475 | $427 | $444,025 |
Academic References (APA 7th Edition)
Peer-reviewed theoretical foundations, clinical diagnostic standards, health economics decision models, and government surveillance data cited in HiddenCare:
05 • How to Cite (APA 7th), Methodology, Research Poster & Sources
Inspect the vector research poster (36×24 in), review complex survey variance proofs, examine formal APA 7th literature citations, and explore author background.
About HiddenCare
HiddenCare investigates how proxy labeling in algorithmic healthcare triage alters who is prioritized for preventive screening under real-world resource constraints.
Decoupling Disease Detection from Healthcare Access
Algorithms trained solely on self-reported doctor diagnoses or administrative billing codes conflate biological disease with healthcare access, misclassifying unvisited sick patients as healthy baseline.
Across N = 4,698 undiagnosed adults from CDC NHANES 2021 to 2023, retraining on observed venous HbA1c swaps 61 individuals at the 20% budget line, uncovering 3 hidden diabetes cases at 3.0x higher precision ($0 extra cost).
Biomarker-augmented triage eliminates an 8.6% sensitivity deficit suffered by uninsured Americans under traditional models, yielding +$25,500 net economic benefit per 10,000 screened individuals.
The Project at a Glance
- The problem
- Predictive screening models frequently train on self-reported clinical history or administrative billing codes (ICD-10). These proxy outcomes measure healthcare access and insurance status rather than pure biological disease, penalizing undiagnosed, marginalized patients at the screening margin.
- What it does
- Re-evaluates N = 4,698 previously undiagnosed US adults from the post-pandemic CDC NHANES 2021 to 2023 cycle under fixed screening budgets (10%, 20%, 30%). Identifies the exact individuals who shift in the triage queue and audits who gets left behind across insurance and income strata.
- How
- Trains regularized Logistic Regression on pre-pandemic NHANES (2017 to March 2020; N = 7,761) comparing Diagnosis-Only labels (Label A) against Biomarker-Augmented labels (Label B: Diagnosis OR HbA1c ≥ 6.5%), holding non-invasive predictors (age, sex, BMI, waist circumference) completely identical.
- Built with
- Python, DuckDB, Parquet, scikit-learn, Gradio, Plotly, pytest, with raw data from the CDC National Center for Health Statistics (NHANES).
Why This Way
- Biological reality over administrative proxies
- Receiving a diabetes diagnosis requires visiting a doctor, having insurance, and affording lab tests. Having elevated blood sugar requires none of those. By training on observed laboratory HbA1c, we decouple biological pathology from socioeconomic access.
- The whole capacity curve
- Public health outreach operates under finite testing capacity (e.g., funding for 1,000 HbA1c kits), not theoretical 0.5 probability thresholds. We evaluate models across continuous budget allocations from 5% to 50% of the population.
- Checked, not just shown
- Every model prediction is validated against observed venous blood laboratory HbA1c from CDC mobile examination centers. Disagreements are verified patient-by-patient at the individual level.
- Readable without an epidemiology degree
- The interface translates complex statistical trade-offs into plain language, interactive capacity curves, and explicit demographic ledgers.
- Honest about sample bounds
- The uninsured evaluation cohort contains N = 422 adults with 13 true cases. We report exact participant counts and avoid overbroad claims regarding algorithmic fairness.
- An engineering project
- Behind this site is an automated, verified data pipeline: binary SAS XPT headers validation, survey weight mapping across cycles, zero target leakage assertion, and pre-computed columnar Parquet lookup tables.
Core Research Questions
- Does training an algorithm on laboratory-augmented biological ground truth alter which undiagnosed adults are prioritized for screening under fixed budget caps?
- Under a 20% outreach budget constraint (K = 939 slots), how many additional unrecognized cases of diabetes are captured?
- What are the clinical, demographic, and socioeconomic profiles of the 61 individuals swapped at the capacity margin?
- Are uninsured and low-income individuals systematically penalized by models trained solely on self-reported clinical history?
- What is the Net Monetary Benefit (NMB) of biological label retraining when accounting for test kit costs and early intervention savings?
Academic Research Poster
One-page research conference poster synthesizing the core scientific hypothesis, CDC NHANES temporal validation protocol, the 61-patient proof, decision curve analysis, and health equity dividends. Click the poster below to enlarge full-screen.
How to Cite HiddenCare (APA 7th Edition)
If you reference HiddenCare findings, the 61-person boundary swap methodology, or the paired complex survey variance analysis in academic papers, policy briefs, or curriculum materials, please use the following formal citation:
Survey Variance & Model Robustness
Addressing complex multistage survey variance via Taylor Series Linearization and verifying algorithmic fairness across nonlinear machine learning architectures.
Taylor Series Linearization (CDC NHANES 95% Confidence Intervals)
Standard IID formulas underestimate variance in clustered populations. We implement Taylor Series Linearization across masked survey strata and primary sampling units (WTPH2YR weights, df = 15):
| Capacity | Model Policy | Weighted Sensitivity (95% CI) | Weighted PPV (95% CI) |
|---|---|---|---|
| 10% Budget | Model A (Diagnosis Only) | 25.25% [16.1%, 34.4%] | 7.33% [4.0%, 10.7%] |
| 10% Budget | Model B (Biomarker Augmented) | 28.77% [19.2%, 38.3%] | 8.23% [4.8%, 11.7%] |
| 20% Budget | Model A (Diagnosis Only) | 40.08% [28.9%, 51.2%] | 5.79% [3.6%, 8.0%] |
| 20% Budget | Model B (Biomarker Augmented) | 42.47% [28.4%, 56.5%] | 6.02% [3.4%, 8.6%] |
| 30% Budget | Model A (Diagnosis Only) | 53.14% [39.9%, 66.3%] | 5.03% [3.2%, 6.8%] |
| 30% Budget | Model B (Biomarker Augmented) | 58.13% [41.8%, 74.5%] | 5.38% [3.2%, 7.6%] |
Nonlinear Model Robustness (Random Forest & Gradient Boosting)
To ensure the 61-patient proof is not an artifact of linear models, we retrained on ensemble decision trees. Across architectures, Model B consistently preserves higher sensitivity and boosts uninsured detection:
| Algorithm | Model A Cases | Model B Cases | Margin Swaps | Swap Precision Ratio | Uninsured Sensitivity Boost |
|---|---|---|---|---|---|
| Logistic Regression | 53 / 116 | 55 / 116 | 61 | 3.0x | +7.7% (+1 case) |
| Random Forest | 60 / 116 | 61 / 116 | 77 | 1.25x | +15.38% (46.2% vs 30.8%) |
| Gradient Boosting | 59 / 116 | 59 / 116 | 132 | 1.0x | 30.8% vs 30.8% |
Author
Cao Thien An Nguyen
Data Science & Economics, University of Wisconsin, Madison
From Ho Chi Minh City, Vietnam, now in Madison, where the winters are long enough to build a whole data pipeline indoors. I blend economics with data science to turn raw data into stories that anyone can follow, and I only trust a number after checking when it was first published.
Fun fact Runs on crispy fried noodles (mì xào giòn): one plate a week through all of 2025, about 52 plates. Revisions to this preference since the first print: zero.
Frequently Asked Questions
- What is "proxy outcome bias"?
- When an algorithm is trained to predict an observable administrative label (such as a billing code or self-reported diagnosis) rather than the true underlying condition. Because receiving a diagnosis requires healthcare access, the model learns predictors of healthcare utilization rather than biological risk.
- Why only 4 non-invasive predictors?
- High-dimensional clinical models requiring laboratory panels cannot be deployed in resource-limited community outreach because they require the very clinical visit we are trying to triage. Non-invasive markers (age, sex, BMI, waist circumference) allow zero-cost community screening.
- Is an elevated HbA1c a definitive diagnosis?
- No. In cross-sectional population surveys like NHANES, a single HbA1c measurement above 6.5% serves as an epidemiological screening criterion. Clinical diagnostic confirmation requires repeat laboratory testing.
- Can I reproduce this entire study?
- Yes. All raw CDC SAS transport files, preprocessing pipelines, model checkpoints, unit tests, and pre-computed Parquet tables are fully open-source and reproducible with a single command:
python run.py.
Clinical & Methodological Disclaimer
Not Medical Advice: HiddenCare is an empirical research and educational policy simulation platform investigating algorithmic proxy outcome bias under resource constraints. It is intended strictly for academic, epidemiological, and data science research. The models, risk scores, and simulations presented on this site do not constitute clinical diagnostic tools, medical advice, or personalized care recommendations.
Epidemiological Surveillance vs Clinical Diagnosis: In cross-sectional population surveys such as CDC NHANES, an elevated laboratory Glycohemoglobin (HbA1c ≥ 6.5%) measurement serves as a biological screening surrogate. Under clinical practice guidelines (such as the American Diabetes Association Standards of Care), an authoritative diagnosis of Type 2 Diabetes requires confirmatory repeat testing or clinical adjudication. Individual patient triage decisions should always be directed by licensed medical professionals.
Data Attribution: This research utilizes public microdata from the continuous National Health and Nutrition Examination Survey (NHANES 2017 to 2023) administered by the National Center for Health Statistics (NCHS), Centers for Disease Control and Prevention (CDC). This independent investigation is neither conducted, endorsed, nor certified by the CDC or NCHS.
Academic References (APA 7th Edition)
Primary academic literature, clinical diagnostic criteria, biostatistical survey methodology, and CDC NHANES biomonitoring microdata informing the HiddenCare algorithmic triage investigation.
Interactive CDC NHANES Variable Codebook
original_name | variable_name | release | file_source | unit | role | description |
|---|---|---|---|---|---|---|
RIDAGEYR | weight | post_pandemic | P_DEMO | probability weight | outcome_definition | Ratio of family income to federal poverty guideline grouped into low middle high or missing |
When clinical history masks biological reality: an empirical investigation of proxy outcome bias in algorithmic triage across N = 12,459 CDC NHANES participants.
Explore
- 01 Capacity Allocation Simulator
- 02 The 61-Person Margin Inspector
- 03 Health Equity & Access Audit
- 04 Health Economics & Decision Curves
- 05 About & Methodology
Data Sources
- CDC NCHS NHANES Continuous Cycles 2017 to 2023
- Glycohemoglobin (GHB) Laboratory Biomonitoring
- ADA Clinical Criteria Diagnostic cut-off HbA1c ≥ 6.5%
- Complex Survey Design WTMEC2YR Examination Weights
Educational and epidemiological research only. An elevated HbA1c measurement in cross-sectional surveillance is a clinical screening surrogate, not an autonomous medical diagnosis. Models and simulation outputs are designed to examine algorithmic fairness under resource limits and must not guide clinical decision-making without qualified medical review.
Microdata sourced from the National Center for Health Statistics (NCHS), Centers for Disease Control and Prevention (CDC). This independent analysis is neither endorsed nor certified by the CDC or NCHS.
Academic References (APA 7th Edition)
Peer-reviewed theoretical foundations, clinical diagnostic standards, health economics decision models, and government surveillance data cited in HiddenCare: