QUANTIX BIO · COMPUTATIONAL MEDICINE · CAMBRIDGE, MA

Medicine, computed in the open.

Quantix Bio is an independent computational medicine lab in Cambridge, MA, run by Karanvir Singh. Nineteen studies and eleven full manuscripts, designed and run in July and August 2026 on public research datasets. Each statistical method is implemented directly in numpy so the mathematics can be read and checked, each figure is generated by the code it describes, and one study reports a result that went the wrong way.

READ THE PAPERS → ENTER THE LAB THE QUANTIX PORTFOLIO
19completed studies
11 draftedmanuscripts, reproducible
160k+real records, beats and molecules
1 keptresult that went the wrong way

Stated first, because it is binding. Nothing on this page is FDA cleared, CE marked, or clinically validated, and nothing here is medical advice. These are methods studies on public research datasets, offered for research use only. Every number is computed out of fold on data the model never touched during fitting, and every study carries a limitations section written to survive a hostile reviewer.

00 · THE MANUSCRIPTS

The papers.

Eleven journal draft manuscripts, each backed by experiments in this lab, each regenerable from public data with one command. The findings below are the papers' own abstracts, compressed.

PAPER 1

Post hoc recalibration can make clinical risk models worse

Eight datasets, four strategies, fifty repeated splits per cell, random effects pooling, a miscalibrated contrast learner and a theory driven simulation that independently reproduces the empirical break even of 100 to 200 calibration observations.

 READ → PAPER 2

Evaluating ward early warning systems by lead time per unit of alarm burden

Two alerting systems with identical detection rates differ two fold in false alarms per patient day; the proposed operating curve makes the difference visible where single threshold metrics cannot.

 READ → PAPER 3

When abstention helps: error concentration under selective prediction

Deferring the least confident fifth of cases removes nine tenths of errors on separable tasks and only a third on the hard tasks where clinical models live; bagged ensembles do not close the gap.

 READ → PAPER 4

Sepsis early warning on real ICU data from two hospital systems

The alarm burden framework applied to hourly vitals from thousands of real ICU patients: derived on one hospital system, externally validated on another, with detection and false alarms per patient day reported where single threshold metrics mislead.

 READ → PAPER 5

A dependency free QRS detector on the complete MIT-BIH database

Pan and Tompkins reimplemented in two hundred readable lines and evaluated beat by beat on all 109,494 annotated beats: 98.6 percent sensitivity with a full census of exactly where and why the classical pipeline fails.

 READ → PAPER 6

Discrimination survives, calibration decays: twenty years of NHANES

A diabetes risk score frozen in 2002 and carried through 2018 on 45,412 adults: ranking ability barely moves while calibration error quadruples, tracking rising prevalence at a correlation of 0.96, and a prospective recalibration rule beats both default policies.

 READ → PAPER 7

Ventricular ectopy defeats RR interval screening for atrial fibrillation

The premise of wearable AF screening, measured beat by beat on 108,534 annotated beats: irregularity detects AF almost perfectly in ectopy free rhythm and collapses on the half of real ambulatory windows that contain ectopy.

 READ → PAPER 8

Nine features under the inter patient paradigm: honest ECG beat classification

Trained and tested on disjoint patients, the evaluation most published accuracies avoid: ventricular ectopy is nearly solved by timing and width, while supraventricular beats are rankable but not thresholdable across hearts.

 READ → PAPER 9

Aggregate monitoring masks a sixfold subgroup calibration disparity

The twenty year drift of Paper 6, decomposed by race, sex and age: probabilities stayed honest for young and white participants while becoming six times worse for Mexican American and older adults, invisible in the aggregate.

 READ → PAPER 10

Removing the race column does not remove the race information

On 14,080 NHANES adults, ordinary anthropometrics and labs reconstruct self identified race and ethnicity at AUC up to 0.76, income adds almost nothing, and a race blind risk score still flags groups unequally. Race blindness is a feature list property, not a behavioural one.

 READ → PAPER 11

Covariate shift is the wrong worry: what actually fails across hospitals

Feature families that fingerprint their hospital transport their sepsis signal intact, while measurement frequency, distributionally identical across sites, collapses and inverts. Provenance, not distributional distance, is the transportability audit axis.

 READ →
01 · THE STUDIES

The studies.

Each study runs end to end from raw public data with Python 3.9 and numpy alone. The figures are the direct output of the code, and each study lists its dataset and original citation.

STUDY /01 · CARDIOLOGY · 303 REAL PATIENTS
Data: UCI Heart Disease, Cleveland subset · Detrano et al, Am J Cardiol 1989

Calibration and abstention in a small real cohort.

On the Cleveland Clinic coronary disease cohort, our logistic model reaches an area under the curve of 0.897. The finding that matters is abstention: when the model declines the most uncertain 43 percent of cases and routes them to a clinician, accuracy on the cases it does answer rises from 82 to 95 percent. The clinically honest shape is a model that handles the clear cases and defers the ambiguous ones.

NEGATIVE RESULT, KEPT: recalibration hurt at this sample size, and we say so
0.897AUC · 5 fold CV
82 → 95%accuracy with abstention
n=303Cleveland cohort
{{FIG1}}
Fig 1 · Selective prediction on UCI Cleveland · out of fold
STUDY /02 · ONCOLOGY · 569 REAL SAMPLES
Data: Wisconsin Diagnostic Breast Cancer · Street, Wolberg, Mangasarian 1993

Beyond AUC: does the model actually help anyone?

Discrimination is not benefit. Decision curve analysis asks the clinical question directly: across every threshold a clinician might hold, does acting on the model beat biopsy everyone and biopsy no one? On the Wisconsin breast mass cohort it does, at every threshold plotted. And the model's largest coefficients read exactly as a pathologist would: irregular, large, deeply concave nuclei.

THE REASONING IS INSPECTABLE AND AGREES WITH PATHOLOGY
0.995AUC · out of fold
+5.8net true positives per 100
n=569WDBC cohort
{{FIG2}}
Fig 2 · Decision curve analysis · WDBC
STUDY /03 · SURVIVAL ANALYSIS · 228 REAL PATIENTS
Data: NCCTG lung cancer trial · Loprinzi et al, J Clin Oncol 1994

Cox regression, written out and proven correct.

On the NCCTG advanced lung cancer trial we implemented Kaplan Meier, the log rank test, and Cox proportional hazards from the partial likelihood upward, in plain numpy. The hazard ratios reproduce the published reference values: 0.58 for female sex, 1.59 per point of ECOG performance status. That agreement is the point. An implementation you can read line by line, validated against the field's standard.

HAZARD RATIOS MATCH THE PUBLISHED LITERATURE
0.58HR · female sex
1.59HR · per ECOG point
p=.0016log rank, sex
{{FIG3}}
Fig 3 · Kaplan Meier by sex · NCCTG lung cohort
STUDY /04 · DRUG DISCOVERY · 1,128 MOLECULES
Data: ESOL solubility set · Delaney, J Chem Inf Comput Sci 2004

How much of a landmark model fits in five numbers?

Aqueous solubility gates oral bioavailability, which makes it the first question asked of any drug candidate. Five transparent physicochemical descriptors, in a ridge regression anyone can audit, recover most of Delaney's landmark ESOL model: 1.19 log units of error against the reference model's 0.91, from a baseline of 2.10. The gap localizes precisely what the missing lipophilicity term is worth.

1.19RMSE · 10 fold CV
0.68R squared
5interpretable descriptors
{{FIG4}}
Fig 4 · Measured vs predicted log solubility · ESOL
STUDY /05 · EPIDEMIOLOGY · A REAL OUTBREAK, FULLY OBSERVED
Data: influenza outbreak report · BMJ 1978;1:587

An epidemic, reduced to two parameters.

In January 1978, influenza swept a closed boarding school of 763 boys, and the infirmary counted every bed, every day. A mechanistic SIR model, integrated with our own Runge Kutta solver and fitted by least squares, reproduces the entire fourteen day epidemic within seventeen boys of error on a peak of 298, implying a basic reproduction number of 3.77. Flatten the curve, computed rather than asserted.

3.77basic reproduction number
292 / 298peak · model vs observed
2.26 dinfectious period
{{FIG5}}
Fig 5 · SIR fit · 1978 boarding school influenza
STUDY /06 · PATIENT SAFETY · SYNTHETIC BY DESIGN
Data: generated by the study code · no clinical source

Early warning scores, measured in false alarms per day.

Ward early warning systems fail through alarm fatigue, so we evaluate on the curve that matters at the bedside: hours of warning against false alerts per patient day. On a simulated cohort, where ground truth is exactly known, a trend aware model halves the alarm burden of a standard threshold score at equal warning time. The data is synthetic and says so in its first line; the contribution is the evaluation method, which transfers to real wards.

SYNTHETIC COHORT · METHODOLOGY STUDY ONLY · NO CLINICAL CLAIM
½the false alarms, equal lead time
6.6 hmedian warning at matched burden
{{FIG6}}
Fig 6 · Lead time vs alarm burden · synthetic cohort
02 · THE INSTRUMENTS

The methods.

Nothing is imported and hoped for. Each of these is implemented in readable numpy, tested, and open to inspection by any reviewer, any regulator, any student.

logistic regression · Newton isotonic calibration · PAVA Kaplan Meier + Greenwood log rank test Cox partial likelihood · Breslow Harrell concordance bootstrap confidence intervals ridge regression ROC · reliability · ECE decision curve analysis selective prediction SIR dynamics · RK4 final size equation an SVG plotter of our own

The entire laboratory runs on python3 and numpy alone. No torch, no sklearn, no matplotlib, no cloud. It runs identically on a hospital workstation behind an air gap, which is the same property the wider Quantix Bio suite is built on.

03 · THE DISCIPLINE

How the lab works.

Real data wherever it exists.

Public research datasets first. Simulation only when ground truth must be known exactly, and then declared in the first line of the study.

Out of fold numbers only.

No metric is ever computed on data the model touched during fitting. That includes calibration maps, which are fit on their own held out split.

Negative results are results.

Study 01 reports that a standard technique made things worse, and explains why. Deleting it would have been easy. It stays.

Limitations written for a hostile reviewer.

Every study names its cohort's age, its site effects, and exactly which claims do not transfer. If a sentence cannot survive review, it does not ship.

Never a clinical claim.

Research software, research use only. The distance between a methods study and a medical device is measured in years of validation, and we say so.

04 · WHAT COMES NEXT

What comes next.

NEXT

Real ICU data

Credentialed access to the MIMIC critical care database under a data use agreement, and the calibration and early warning studies rerun on real patients.

THEN

Preprint

The small sample recalibration failure of Study 01, replicated across further public clinical cohorts and posted as a citable preprint.

THEN

Peer review

Faculty collaboration in clinical informatics and biostatistics, and a methods note submitted to an open medical informatics journal.

ALWAYS

Open mathematics

Whatever the lab grows into, every method stays readable, every figure stays reproducible, and every limitation stays printed.

QUANTIX BIO · CAMBRIDGE, MA

Questions and criticism welcome.

If you work in clinical informatics, biostatistics, or medicine and see something wrong or missing in these studies, that is exactly the mail this lab wants.

VIEW FULL STUDIES → CONTACT