An Author-Level Component of Clinical Note Length in MIMIC-III Resident Progress Notes: A Testable Hypothesis for Clinical Natural Language Processing

Background: Clinical natural language processing (NLP) treats notes as records of the patient, but notes are also written by individual clinicians whose habits may shape the text. Objectives: To test, using note length, whether clinical text carries an author-level component after accounting for principal diagnosis. Methods: This retrospective cross-sectional study analyzed 62,682 resident progress notes by 435 physicians in the Medical Information Mart for Intensive Care III (MIMIC-III), a de-identified single-center critical care database (Boston, 2001 to 2012). Study 1 related word count to a fixed 54-term content proxy, overall and at constant proxy value. Study 2 retained notes from the ten most frequent principal diagnoses (20,581 notes; 266 physicians), centered word count on diagnosis means, and estimated the author intraclass correlation coefficient (ICC), with and without an admission component. Results: Word count correlated with concept count (r = 0.695; 95% CI 0.680 to 0.710). At constant concept count (11; n = 5,999), the 90th-to-10th percentile length ratio was 1.93. After diagnosis centering, the pooled ICC was 0.225 (95% CI 0.189 to 0.257). With admission modeled, the author share was 0.187 and the admission share 0.244; retaining one random note per author-admission gave 0.201. Physicians at the 10th and 90th percentiles differed by 334 words per note. Conclusions: The content proxy accounted for approximately half the variance in note length, and resident note length clustered measurably by author after accounting for principal diagnosis, including when repeated notes within admissions were modeled. Care-unit conventions, calendar period, and case mix cannot be excluded. Whether this author component affects models trained or evaluated on clinical notes was not tested; we outline three pathways by which it could, with a test for each, and propose author-aware reporting items.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24233693
Primary Topic
Machine Learning in Healthcare
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

An Author-Level Component of Clinical Note Length in MIMIC-III Resident Progress Notes: A Testable Hypothesis for Clinical Natural Language Processing

Tejus B R
Zenodo (CERN European Organization for Nuclear Research)
Machine Learning in Healthcare
preprint

An Author-Level Component of Clinical Note Length in MIMIC-III Resident Progress Notes: A Testable Hypothesis for Clinical Natural Language Processing

Tejus B R
preprint en

Abstract

Background: Clinical natural language processing (NLP) treats notes as records of the patient, but notes are also written by individual clinicians whose habits may shape the text. Objectives: To test, using note length, whether clinical text carries an author-level component after accounting for principal diagnosis. Methods: This retrospective cross-sectional study analyzed 62,682 resident progress notes by 435 physicians in the Medical Information Mart for Intensive Care III (MIMIC-III), a de-identified single-center critical care database (Boston, 2001 to 2012). Study 1 related word count to a fixed 54-term content proxy, overall and at constant proxy value. Study 2 retained notes from the ten most frequent principal diagnoses (20,581 notes; 266 physicians), centered word count on diagnosis means, and estimated the author intraclass correlation coefficient (ICC), with and without an admission component. Results: Word count correlated with concept count (r = 0.695; 95% CI 0.680 to 0.710). At constant concept count (11; n = 5,999), the 90th-to-10th percentile length ratio was 1.93. After diagnosis centering, the pooled ICC was 0.225 (95% CI 0.189 to 0.257). With admission modeled, the author share was 0.187 and the admission share 0.244; retaining one random note per author-admission gave 0.201. Physicians at the 10th and 90th percentiles differed by 334 words per note. Conclusions: The content proxy accounted for approximately half the variance in note length, and resident note length clustered measurably by author after accounting for principal diagnosis, including when repeated notes within admissions were modeled. Care-unit conventions, calendar period, and case mix cannot be excluded. Whether this author component affects models trained or evaluated on clinical notes was not tested; we outline three pathways by which it could, with a test for each, and propose author-aware reporting items.

Zenodo (CERN European Organization for Nuclear Research)
Machine Learning in Healthcare
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon