Letters to the Editor - In Response To - OphthoACR (Ophthalmology Automated Chart Review) - Request for Deployment-Ready Scalability Clarification

Sep 28, 2026

Uploaded by awa

Letter to the editor responding to OphthoACR’s performance claim that cohort processing runs in O(1) time for AI review versus O(n) for manual review. The author argues that practical wall-clock time depends on achievable parallelism across EHR export, OCR, network services, and LLM API quotas, plus an inherently serial portion governed by Amdahl’s law. Queueing effects under rate limits and network constraints further shape throughput. The letter requests a deployment-ready scalability profile (latency, throughput vs concurrency, measured serial fraction, and operating points with constraints and cost).

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Response and performance-claim critique

Dependence on parallelism and serial components

Queueing effects under constraints

Recommended scalability reporting

Latency and throughput under controlled concurrency

Measured serial fraction and operating points

Letters to the Editor In Response To: OphthoACR (Ophthalmology Automated Chart Review): An AI-Powered Tool for Complete Automation of Ophthalmology Chart Reviews and Cohort Data Analysis Chen et al. present Ophthalmology Automated Chart Review (OphthoACR), an ambitious, end-toend pipeline that turns unstructured ophthalmology charts into structured, analyzable data with accuracy and speed.1 The work is timely and thoughtfully engineered. One aspect, however, would benefit from clarification because it directly affects realworld deployment and readers’ expectations of scalability: the assertion that cohort processing “results in a total runtime of O(1) in artificial intelligence (AI) review as opposed to O(n) in manual review.” In theory, the claim that parallel processing breaks the strictly linear dependence of review time on cohort size is correct. In practice, however, wallclock time for automated review is determined by (1) the effective parallelism that can actually be sustained (through the electronic health record [EHR] export pipeline, optical character recognition [OCR], network, tokenizer/embedding services, and large language model application programming interface [API] quotas), and (2) the fraction of the pipeline that is inherently serial (data staging, deidentification, quality checks, and post-processing) . Amdahl’s law reminds us that even a small serial fraction places a hard upper bound on achievable speed-up, regardless of how many workers are added; strictly speaking, “O(1)” runtime assumes unbounded, cost-free parallelism and negligible serial work.2 Moreover, when queues form — commonly under API rate limits, batch scheduling, or hospital virtual private network constraints — throughput is governed by basic queuing behavior: how many charts per hour the system can actually process (service rate) versus how quickly new charts are submitted (arrival rate), rather than by nominal parallelism alone.3 To strengthen the translational impact of Ophtho ACR and make its performance claims portable across institutions, I suggest reporting a simple, deploymentready scalability profile alongside accuracy metrics: • Single-worker latency (median and 95th percentile) per chart and per source document; • Throughput (charts/hour) as a function of controlled concurrency (e.g. P = 1, 4, 8, 16, and 32) using the same hardware, network environment, and API quotas; • The measured serial fraction (time in EHR export, OCR, de-identification, and post-processing) to provide an Amdahl-law upper bound on speed-up; • Operating points under institutional constraints, including (a) achieved service rate and arrival rate under real API and network limits and (b) corresponding cost/compute intensity per chart at those operating points. For clinicians, these quantities translate directly into practical questions: whether a 10,000-patient chart review finishes over a weekend or over a month, and what resources are required to support that workload. Including such a scalability profile would not detract from the authors’ core advance; rather, it would convert an excellent proof-of-concept into an operational blueprint that health systems can use to plan capacity, budget responsibly, and reproduce results. By aligning performance claims with established scaling principles and reporting them in deployment-relevant terms, readers will be better able to anticipate time-toresult under local constraints and, ultimately, accelerate responsible adoption. Henry Bair 1 1 Wills Eye Hospital, Philadelphia, PA, USA. e-mail: [henry.c.bair@gmail.com](henry.c.bair@gmail.com) Copyright 2025 The Authors [tvst.arvojournals.org | ISSN: 2164-2591](tvst.arvojournals.org | ISSN: 2164-2591) This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4 .0 International License. 1 Letters TVST | December 2025 | Vol. 14 | No. 12 | Article 32 | 2 References 1. Chen KM, Chen KW, Diaconita V, Chang

FAQs

Why does the author question the O(1) runtime claim for cohort processing?

Because real wall-clock time depends on effective parallelism across the pipeline and on a serial fraction that limits speed-up per Amdahl’s law. Queueing under API rate limits and network constraints can also govern throughput.

What practical factors determine automated review throughput beyond nominal parallelism?

Throughput is determined by service rate (how quickly the pipeline processes charts) versus arrival rate (how quickly charts are submitted), which becomes especially relevant when queues form under operational limits.

What scalability metrics does the author recommend to make OphthoACR’s performance portable across institutions?

A deployment-ready scalability profile including single-worker latency (median and 95th percentile), throughput (charts/hour) versus controlled concurrency, the measured serial fraction across pipeline stages, and operating points under institutional constraints with corresponding cost/compute intensity.

Sep 28, 2026

Uploaded by awa