◇FeatureGraph

Research

A guide to FeatureGraph research.

FeatureGraph is an open research program on making scientific time-series analysis inspectable and reproducible. Each study states what its source data and methods provide, what FeatureGraph automated, and which conclusions still need scientific judgment. The research and its methods are open science.

Each study below used analysis code written before the FeatureGraph compiler could check every step of an analysis. Future studies will run through the compiler.

Representation

What is an object?

How a stream of measurements becomes bounded, inspectable objects with measurements and relations.

Completed · earlier analysis code

BIDMC respiration workflow

Question. Can a declared signal construction be applied across a whole cohort while keeping the observations, states, boundaries, measurements and evidence behind every object?

Result. On subject 1, all 169 objects from an independent reference analysis were matched, with a median peak position difference of 10 samples. Across all 53 subjects, 7,086 objects matched the reference, 1,047 were found only by FeatureGraph, and 82 only by the reference. The 1,047 are kept as bounded episodes, not called errors. Temporal symmetry did not agree between the two methods, because they place trough boundaries differently. This is not clinical validation.

Read the study ↗Open the notebook ↗Read the framework paper draft ↗

Demonstration · earlier analysis code

Accumulation as a dependent object

Question. Can one object be built from another, keeping the parent’s identity and completeness?

Result. Within each candidate respiration cycle, an accumulation object records the area above an explicit baseline, with its timing and centroid. It is waveform area, not inhaled volume, unless the sensor has been calibrated for that meaning.

See the accumulation curvesRead the update

Method

Analysis choices are part of the specification

Why preprocessing decisions belong in the analysis specification and have to be stated.

Manuscript · earlier analysis code

Smoothing is part of the analysis specification

Question. Is smoothing a neutral cleaning step, or a choice that changes what the analysis says exists?

Result. On the 53 BIDMC recordings, the same construction run with a 1-sample window and with a 100-sample window gives object counts that correlate at only 0.39. The ratio between the two counts runs from 1.14 to over 40, depending on the recording.

Read the manuscript ↗

Transfer

Across data and domains

Whether a construction chosen on one dataset holds on others without being changed.

Completed · earlier analysis code

Tennessee Eastman reactor-pressure transfer

Question. Does a construction chosen on one run hold unchanged on held-out runs, normal operation and other fault classes?

Result. The construction was chosen on Fault 2 run 10 and applied unchanged to nine held-out Fault 2 runs and ten normal-operation windows. Peak size separated all ten Fault 2 runs from all ten normal windows. Prominence and event count did not. Some other fault classes produced equal or larger peaks, so this is a repeatable pressure representation, not a Fault 2 classifier.

Read the study ↗Open the notebook ↗

Completed · earlier analysis code

PhysioNet wearable stress protocols

Question. Can the stress-protocol stages and button-marked boundaries declared by the dataset’s authors be preserved exactly, with one schema across both protocol versions?

Result. All 248 declared protocol occurrences from 33 participants kept both source boundaries exactly. All 248 were joined to the matching self-report, and 99 of 99 validation checks passed. The study does not detect stress or validate a physiological biomarker.

Read the study ↗

Interoperability

With other methods

Whether the object representation works with detectors and models built by others, without taking over their role.

Completed · earlier analysis code

CLaP state detection

Question. Can the states an independent detector found be turned into bounded objects without changing the detector’s labels?

Result. CLaP found three recurring states and eight change points in a 20,700-sample benchmark series. FeatureGraph represented them as nine bounded objects with eight relations, and reconstructed all 20,700 labels exactly. The study tests the representation of CLaP’s output, not CLaP’s detection quality.

Read the study ↗Open the notebook ↗

Exploratory demonstration · earlier analysis code

Nori on PhysioNet protocol objects

Question. Can an independent tabular foundation model run directly on FeatureGraph protocol objects?

Result. Yes, without a study-specific adapter. As a predictor of self-reported stress, nothing beat a plain training-mean control, and every learned model had negative out-of-fold R². The demonstration does not show that these summaries predict stress for unseen participants.

Read the study ↗

Checked agents

Language models asking checked questions

Whether a language model can ask questions of the data while a separate program checks and runs them.

Recorded demonstration · checked queries

The agent demo

Question. Can a language model put a question to a table without seeing the data or doing any arithmetic?

Result. In the demo, a model turns a plain-English question into a short query. A separate program checks the query, runs it, and saves a receipt for every run. It is a recording of 13 runs on a small invented table, and it does not measure accuracy.

Open the demo

How to read it. Each row of the demo is one run. Green rows were answered. Red rows were stopped by a check before anything ran, which shows what the checks catch. Purple rows were refused because the answer was not unique, and amber rows were declined by the model.

More

Where the rest of the record lives.