The Research Radar, Issue 001: Ten Digital Health and Clinical AI Preprints from 1 to 15 November 2026, Plus a Personal Health Records Spotlight

The Research Radar · 15 November 2026 · Issue 001 An alternate-day briefing on the ten highest-ranked recent preprints in digital health and clinical AI, plus a personal health records spotlight, written for India’s health-IT community. Preprints are early work. We say what the authors report, and where it stops.

In this issue

• One benchmark (NPCBench) reports a composition gap: strong performance on individual tasks does not ensure jointly correct diagnostic and care answers for the same patient.
• Three EHR papers take three angles: what the record actually encodes, how clinicians used a record-search assistant in one health system, and how a language model can be adapted to EHR event text.
• TransXplainRRG and Mixed Signals flag their own limits. TransXplainRRG describes its radiologist evaluation of 100 reports as preliminary, and the Mixed Signals authors say their approach needs further investigation before practical use.
• A personal health records spotlight: three preprints on PHRs and health apps, and three live PHR or health-record apps, one each from India (Aarogya Setu 2.0), the Middle East (Sehhaty, Saudi Arabia) and the US (Google Health app), each described from its operator’s or government’s own pages.

Updated 15 November 2026. The method note that opened this issue now sits at the end, under “How this issue was made”, and the personal health records spotlight is new. The patterns section is unchanged, and the ten papers are unchanged apart from one caveat wording fix (Neuro-VL). The front matter, tagline, preprint notice, numbering note, FAQ, “In this issue” and sign-off were updated to match. The method section also gains a process version line, the personal health records track method and a description of how the issue was checked.

Preprint notice. These are arXiv preprints. They have not necessarily been peer reviewed, and results may change. We report what the authors state in their abstracts and, where marked, in their full texts. None of this is clinical guidance.

The ten papers

The papers are numbered by shortlist rank, 2 to 11, because one candidate was skipped on reading (see “How this issue was made” at the end).

2. SkinMind: Equipping Agentic RAG with Human-Inspired Graph Probing for Reliable Dermatology AI

Jing Hao, Weinan Zhou, Zhihong Lu et al. · arXiv:2611.08935 · 6 November 2026 · Rank basis: Hugging Face upvotes (4)

• What it does: an agentic retrieval system in which an orchestrator probes a dermatology knowledge graph with eight tools to find text and image evidence. The authors built the graph from a 62M-token corpus of dermatology guidelines and textbooks and 423K images, with over 1.0M nodes and 2.0M edges.
• What it reports: across six dermatology visual question answering benchmarks, SkinMind improves the average score over direct answering by 8.70 points with GPT-5.4-mini and 8.89 points with Gemini-2.5-Flash, and outperforms competing retrieval methods.
• Caveat: results are on benchmarks, not in clinical use. Preprint, not necessarily peer reviewed.
• Why it could matter in India (our reading): guideline-grounded retrieval with visible evidence is a pattern worth testing for specialty teleconsultation.

3. NPCBench: A Clinical Apprenticeship Benchmark for Multimodal Care-Pathway Reasoning in Nasopharyngeal Carcinoma

Pengkai Wang, Wei-Wei Zhang, Yan Li et al. · arXiv:2611.08369 · 5 November 2026 · Rank basis: Hugging Face upvotes (3)

• What it does: a benchmark for nasopharyngeal carcinoma that follows specialist training, from guidelines to case planning, longitudinal review and consultation. It has 4,777 component and consultation items and a 100-patient cohort with 417 scored diagnostic and care tasks.
• What it reports: across 30 evaluated models, strong performance on individual tasks does not ensure jointly correct diagnostic and care answers for the same patient, which the authors call a composition gap. Among 21 multimodal models, strict care-pathway success reaches at most 23%.
• Caveat: one cancer type, and a benchmark rather than a clinical trial. The arXiv comments say it was accepted at the NeurIPS 2026 Evaluations and Datasets Track; we read only the preprint abstract.
• Why it could matter in India (our reading): when evaluating any clinical AI, test whole pathways for one patient, not only isolated questions.

4. Mixed Signals: Weight-Space Programming for EEG and ECG

Dara Varam, Taraneh Aminosharieh Najafi, David Atienza · arXiv:2611.13758 · 10 November 2026 · Rank basis: Hugging Face upvotes (3)

• What it does: lets a language model read and edit the weights of small networks trained on EEG and ECG data, in effect “training” by reasoning instead of gradient descent.
• What it reports: frontier models debugged faults planted in biosignal models without being told what they were, and could write a network’s weights from scratch given the architecture, task and a few training subjects. In one set of experiments a language model matched gradient-descent training with 6x fewer training subjects. Networks written by Claude Opus 5.5 scored higher than classic hand-crafted EEG pipelines when built from the same 5 subjects.
• Caveat: the authors say the approach cannot improve already well-trained classifiers, cannot reliably recognise a network’s task from its weights, and still needs further investigation before practical use. Preprint.
• Why it could matter in India (our reading): small models on wearables and edge devices are a plausible fit for low-resource settings, but this is early research.

5. PhysioWave-T: Token-Efficient Wavelet Foundation Models for Biosignals

Yanlong Chen, Simone Benatti, Luca Benini et al. · arXiv:2611.08969 · 6 November 2026 · Rank basis: Hugging Face upvotes (2)

• What it does: a family of foundation models for EEG, ECG and sEMG signals that compresses the token sequence with a technique the authors call scale folding. They pre-trained a 30M-parameter EEG model on over 15M windows, alongside ECG (30M) and sEMG (12M) encoders.
• What it reports: the authors report that, across four EEG routes, scale folding cuts transformer latency by 4.0 to 7.2x and peak memory by 6.8 to 12.7x over scale expansion, without accuracy loss. The EEG encoder leads 8 of 9 metrics against BENDR, BIOT, LaBraM and EEGPT. On SeizeIT2, their fusion method raises event sensitivity at 11 false alarms per hour from 61.4% to 74.4% over the best unimodal baseline. Code and models are released.
• Caveat: the abstract describes benchmark results, not a clinical deployment. Preprint.
• Why it could matter in India (our reading): efficiency gains matter most where compute and connectivity are limited.

6. Every Pixel Is an Option: A System One Model for Medical Ultrasound Biometry

Moein Heidari, Junbo Rao, Yudan Chen et al. · arXiv:2611.03360 · 2 November 2026 · Rank basis: Hugging Face upvotes (2)

• What it does: Kev-Bio, a vision-language model that uses a clinical measurement question to guide landmark detection in ultrasound images, producing spatial probability maps from which caliper positions and measurements are derived.
• What it reports: on the nine-view FU/GU Biometry 2026 dataset, a mean radial error of 22.50 pixels and a clinical measurement mean absolute error of 15.38, outperforming the compared baselines overall. Code and models are available.
• Caveat: the abstract does not state the units of the measurement error or compare against operator variability. Preprint.
• Why it could matter in India (our reading): automating caliper placement could reduce interoperator variability in scans, if it holds up clinically.

7. A frontier machine intelligence for human neuroimaging

Chenhui Zhao, Akshay Rao, Akhil Kondepudi et al. · arXiv:2611.15094 · 11 November 2026 · Rank basis: Hugging Face upvotes (1)

• What it does: Neuro-VL, a vision-language model trained on nearly one million neuroimaging studies from routine clinical care, which the authors say can interpret complete examinations and discuss findings in natural language.
• What it reports: across internal and international datasets, the authors say it outperformed frontier general-purpose and specialised medical models at diagnosing, localising and describing abnormalities and at generating radiology reports. They also report that it compared serial examinations to assess progression and more reliably resisted misleading clinical assumptions than leading frontier models from OpenAI and Google.
• Caveat: these are the authors’ own comparisons, and the abstract gives no performance figures. Preprint.
• Why it could matter in India (our reading): the training data came from health systems’ routine care, which raises the question of what comparable Indian multi-hospital data collaborations would need.

8. The Dark Matter of Clinical Data: An Identification Limit on Disentangling Physiology from Care Process in electronic health records

Mohammad Shahin, Krittaphas Chaisutyakorn, Dimitrios Proios et al. · arXiv:2611.11726 · 9 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

• What it does: a formal analysis of how EHR representations mix patient physiology with the clinical process that decides what is measured, when and for whom.
• What it reports: the authors show that no division of the information carried by laboratory value and ordering channels is identified from a single dataset, and that cross-site invariance is a falsifiable implication rather than an identification strategy. A worked example gives a 9.6 percentage point cross-site gap under an unchanged biological mechanism. They recommend reporting cross-site analyses as tests of a stated joint hypothesis.
• Caveat: a theoretical result, with a worked example rather than a clinical validation. Preprint.
• Why it could matter in India (our reading): anyone pooling records across hospitals should expect that differences in ordering practice, not only in patients, can shape what a model learns.

9. Real-World Adoption, Use, and Governance of an LLM-Based Electronic Health Record Search and Synthesis Platform

Jiunn-Tyng Yeh, Michael Gao, Marshall Nichols et al. · arXiv:2611.08672 · 5 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

• What it does: a longitudinal study of Scout, an EHR-integrated LLM platform that gives cited search and synthesis of patient records, during staged deployment across a multi-hospital academic health system.
• What it reports: over 11 months, 2,635 users completed 80,465 tasks. 1,768 clinical users started within 29 days of launch, which included required training, and daily initiation rose 21.2-fold at launch, from 2.9 to 61.0 users per day. Of care practitioners who completed the required course, 77.9% went on to use it, and the top user quintile accounted for 59.0% of tasks. The authors suggest rapid, broad adoption needs clinician-led task design and evaluation, governance proportional to reach, centralised capacity planning and support, and training integrated with access.
• Caveat: one health system, and the abstract reports adoption and use, not accuracy or patient outcomes. Preprint.
• Why it could matter in India (our reading): hospitals piloting record-search assistants could borrow the metrics (adoption after training, concentration of use) as a template for what to measure.

10. Evaluating Pretrained Language Semantics for Structured EHR Prediction Tasks

Ashvin Gupta, Brendan Delaney, Alessandra Russo · arXiv:2611.17053 · 12 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

• What it does: tEHR, which represents each EHR event by its natural-language description, continues pretraining a causal language model on that text, then fine-tunes it with LoRA as a longitudinal patient encoder.
• What it reports: the authors evaluate tEHR on four prediction tasks on MIMIC-IV and on pancreatic cancer prediction on CPRD at 1 to 12 month horizons, with the Mistral-7B-v0.3 backbone performing best overall. Performance declined when clinical descriptions were removed or misaligned, when numerical values were omitted, or when either adaptation stage was removed. Explicit elapsed-time descriptions added little.
• Caveat: two named datasets, and the abstract gives no clinical deployment. The arXiv comments say it was accepted for publication in Proceedings of Machine Learning Research, volume 346; we read only the preprint abstract.
• Why it could matter in India (our reading): using concept names rather than local codes may ease transfer between hospitals with different coding habits, which is a question to test, not a finding here.

11. TransXplainRRG: A Radiologist-Evaluated LVLM for Chest Radiograph Report Generation

Chayan Mondal, Duc-Son Pham, Tele Tan et al. · arXiv:2611.16393 · 12 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

• What it does: chest X-ray report generation that adds patient medical history through a text encoder and refines reports by extracting abnormal findings, aiming for both performance and explainability.
• What it reports: trained and tested on MIMIC-CXR and further evaluated on the unseen IU X-Ray dataset, it scored 0.327 and 0.605 on the GREEN metric respectively, both higher than MAIRA-2 under the same setting. The authors also report a preliminary evaluation of 100 generated reports by radiologists.
• Caveat: the radiologist evaluation is described as preliminary. Preprint.
• Why it could matter in India (our reading): radiology report drafting is a likely early use of vision-language models, with the reporting radiologist remaining the author of record.

Patterns across the ten

• One benchmark tests the whole pathway. NPCBench reports strict care-pathway success of at most 23% among 21 multimodal models, even where single tasks look strong.
• Three EHR papers take three angles. One analyses what the record encodes (the identification paper), one studies how clinicians use a record tool (Scout), and one adapts a language model to EHR event text (tEHR).
• Wearables appear in both biosignal papers, and edge devices in one. PhysioWave-T reports latency and memory reductions and a wearable seizure-detection result. The Mixed Signals authors say weight-space programming can be useful for small networks such as those on edge devices and wearables, but needs further investigation.

Personal health records spotlight

Personal health records (PHRs) and the apps that carry them are where patients meet their own health data. This section pairs three recent preprints with three live PHR or health-record apps, one each from India, the Middle East and the US, the three markets we are tracking in the near term. The apps are described from their operators’ or governments’ own pages. We have not tested them.

Three preprints on PHRs and health apps

These three came from a separate shortlist (see “How this issue was made”). None was featured on Hugging Face, so all were ranked by keyword relevance.

A prospective randomized study of patient-centered AI for clinical visit preparation

Xincheng Tan, Matthew Thompson, Xiang Ji et al. · arXiv:2610.41437 · 30 October 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

• What it does: an LLM-based visit preparation agent that guides patients before an appointment by gathering clinical context, setting visit expectations, providing relevant health information, building a prioritised list of questions for their clinician and generating a summary.
• What it reports: in a prospective consented randomized study in the Google Health app, 802 U.S. adults held 866 health-related conversations with the agent (standard or concise variant) or a minimally prompted control chatbot built on the same LLM. In clinician reviews of completed sessions (n = 272), comprehensive preparation was 67.2% against 50.6% (P < 0.05) and alignment with patients’ health context was 98.9% against 90.4% (P < 0.01). In patient post-surveys (n = 460), visit preparedness was 73.8% against 57.0% (P < 0.01). The authors report high medical safety and appropriateness across all study arms, and that clinician and participant metrics were not significantly correlated.
• Caveat: per the full text, participants were English-speaking U.S. adults using the Google Health app on an Android phone, 42.3% of included conversations ended before completion, and clinical evaluation covered completed sessions only. The abstract reports no clinical outcomes, and the authors say future studies should look at in-visit communication and clinical outcomes. Preprint.
• Why it could matter in India (our reading): a pre-visit workflow that ends in a question list for the clinician is a pattern that ABDM-connected PHR apps could test with their own users.

Bad Diagnosis: Analyzing Consent and Data Collection Practices of Mobile Medical Apps

Ethan Myers, Lorenzo De Carli, Shantanu Sharma et al. · arXiv:2610.23851 · 16 October 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

• What it does: a measurement study of consent and data collection practices of mHealth apps on the Google Play Store, 10,000+ apps in the dataset, using static analysis and privacy policy examination and, per the full text, dynamic analysis of apps as they run, with an operational definition that separates health apps from general wellness programs.
• What it reports: the authors report that 69.15% of apps request at least one permission unrelated to their stated purpose (per the full text, permissions were analysed for 10,036 apps). Over 52.9% request location data, and 50.5% and 34.9% request camera and microphone access, despite no functional justification based on app descriptions and stated purposes. Per the full text, the consent figures come from the 3,599 apps (34% of the dataset) that could be launched and analysed: 253 of them (7%) displayed any consent mechanism, so nearly 93% showed none, which the authors say directly violates regulatory requirements such as GDPR. Among the apps that did obtain consent, 1 in 14 show a gap between the consent obtained and the personally identifiable information collected. The authors plan to release their dataset and measurement tools.
• Caveat: Google Play apps only. Per the full text, the method observes only initial app execution before account creation, and the authors say app failures introduce selection bias. Preprint.
• Why it could matter in India (our reading): PIB says every new app must be tested in a sandbox and pass a security audit that checks how the app handles and protects patient data before it can connect to ABDM. This paper does not study ABDM apps, but it examines the kind of permission and consent practices such checks bear on.

From Structured Records to Patient Narratives: Preserving Clinical Alignment in LLM-Generated Health Summaries

Bayram Ayadi, Florian Geissler, Johanna Schmidhuber · arXiv:2610.31467 · 22 October 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

• What it does: investigates evidence-grounded health summarisation for menstrual health monitoring. LLMs turn structured records into patient-accessible narratives, using a constrained generation framework that combines structured evidence conditioning with retrieval-augmented generation. The authors motivate this by wearable and digital health systems generating longitudinal physiological records.
• What it reports: evaluated with natural language inference, atomic factuality and readability measures across three LLMs and 4,500 generated summaries, explicit grounding constraints consistently improve clinical alignment, particularly uncertainty preservation. Effects on omission are model-dependent, and the improved alignment comes at the cost of increased reading difficulty and medical terminology.
• Caveat: one use case, menstrual health monitoring. Per the full text, the dataset is semi-synthetic, generated by the authors from a causal model, and the authors say no real patient records or wearable device data were used, neither the generated records nor the ground truth were reviewed by a clinician, and the findings should not be taken as validated on real-world data distributions. Preprint.
• Why it could matter in India (our reading): the trade-off the authors report between clinical alignment and readability is worth testing in any PHR feature that rewrites records in plain language.

Three PHR and health-record apps in the wild

These are not ranked and are not a popularity measure. Each is a PHR or health-record app from one of our three tracked markets, with an update or announcement on the operator’s or government’s own pages in 2026.

Aarogya Setu 2.0: a national health app, India

Sources: PIB’s release of 29 June 2026 and its ABDM document of 6 July 2026.

• What it is: PIB says Aarogya Setu, developed during COVID-19 for contact tracing, was handed to the National Health Authority and revamped as a personal health record and citizen-facing digital health application under the Ayushman Bharat Digital Mission. Aarogya Setu 2.0 was launched on 29 June 2026 by the Union Health Minister, and PIB cites a user base of nearly 20 crore downloads.
• Features, as PIB lists them: ABHA creation, health record management and Scan & Register. AI to digitise uploaded records such as lab reports, including Google’s open-source Medical Data Toolkit to turn scanned PDFs and images into machine-readable data that feeds a personalised health dashboard. Vitals, reminders, goal tracking and wearable syncing for steps, calories, heart rate and glucose. PM-JAY wallet and private insurance details, nearby facilities and real-time blood unit availability through e-RaktKosh, and family management. PIB’s ABDM document adds teleconsultations and in-person appointments booked through UHI.
• Scale and consent context (PIB, 6 July 2026): over 104 crore health records linked to over 93 crore ABHA accounts. PIB says records stay with whoever created them and are shared within the ABDM network only when the patient consents, and that its Scan and Share service cut waiting times from about one hour to 2 to 5 minutes in an IIHMR study, with over 23.21 crore ABHA-linked tokens issued as of 18 June 2026. These are government-reported figures.
• Our reading: like Google, India’s PHR builds on an app that already had users, and PIB positions it as an entry point to ABDM services.

Sehhaty: a ministry app, Saudi Arabia (Middle East)

Sources: the Saudi Ministry of Health’s news item of 9 August 2026 and its announcements of 15 August 2026 and 10 June 2026.

• What it is: the Ministry says virtual care extends to beneficiaries at home through the Sehhaty app, which exceeded 31 million users in 2025. The Ministry’s pages do not call Sehhaty a personal health record. They describe prescription viewing and a smart assistant integrated with the beneficiary’s health record.
• Capabilities, as the Ministry describes them (9 August 2026): appointment booking, viewing prescriptions, searching for medications and issuing sick leaves. It also incorporates AI technologies: a digital twin that analyses health status and gives personalised preventive recommendations, and a smart assistant that helps interpret medical results and provides medication and appointment reminders through integration with the beneficiary’s health record. The Ministry also says the Smart Health Coach, launched in collaboration with Google through the Sehhaty app at the Global Health Exhibition 2025, gives personalised health guidance and preventive recommendations based on beneficiaries’ data and health indicators.
• How it is being extended: on 10 June 2026 the Ministry launched an e-consultation to enhance medical report services in the app, covering residency, weapons licensing and driving fitness medical examination reports. On 15 August 2026 it launched an e-consultation on proposed features, including appointment attendance confirmation and a child autism screening questionnaire, or both. These pages describe proposed features, not shipped ones.
• Our reading: the Ministry opened e-consultations on proposed features on 10 June and 15 August 2026, and the app carries transactional services such as sick leaves alongside its AI features.

Google Health app: a consumer platform, United States (medical records sync)

Sources: Google’s posts of 7 May 2026 on the Google Health app and the Google Health Coach.

• What it is: the Fitbit app becoming the Google Health app, which Google describes as a centralized place for health and wellness data. It starts rolling out to all users on 19 May 2026 as an update to existing Fitbit users, with four tabs: Today, Fitness, Sleep and Health. Google describes the app as rolling out to all users and the Coach as globally available. The medical records sync it describes is for the US, which is why this is our US entry.
• Capabilities, as Google describes them: it works with hundreds of apps and devices through Health Connect, Apple Health or the Google Health APIs. In the US, users can sync medical records to view lab results, vitals and medications and how they change over time, with records stored in the app and under the user’s control. The Google Health Coach, built with Gemini, is available through a Google Health Premium subscription and can give summaries of medical records on the Health tab. Google says it committed to not using Fitbit user health and wellness data for Google Ads.
• Our reading: Google updated an existing app rather than launching a new one. Records sync is described for the US, and the AI coach sits behind a subscription.

Across the three apps and three papers (our reading)

• Google and India’s Aarogya Setu 2.0 both build on an app that already had users: Google updated the existing Fitbit app, and PIB cites nearly 20 crore downloads for Aarogya Setu.
• Each app’s source pages mention AI features: Google’s Coach is a Premium subscription feature, Aarogya Setu 2.0 lists AI digitisation of uploaded records among its features, and the Saudi Ministry describes Sehhaty’s digital twin and smart assistant.
• Google appears in all three sources: it operates the Google Health app, PIB says Aarogya Setu 2.0’s AI document digitisation uses Google’s open-source Medical Data Toolkit, and the Saudi Ministry says the Sehhaty Smart Health Coach was launched in collaboration with Google.
• The three preprints each address a different question relevant to patient-facing health data and apps: whether permissions and consent match purpose (mHealth apps on Google Play), whether a structured workflow helps patients prepare for a visit (in the Google Health app), and whether LLM-written summaries of structured health records stay faithful.

The takeaway

Read these as a sample of recent preprints chosen by a stated method, not as evidence of what works in a ward. The most actionable item for a hospital IT team this fortnight is probably the adoption data from a real deployment. The most sobering is one benchmark (NPCBench) reporting that strict care-pathway success reached at most 23% among 21 multimodal models. We will repeat this on alternate days with a fresh window.

How this issue was made

Process version 2, 15 November 2026. Version 1 put the method note at the top of the issue and had no personal health records section. This section will record future changes to how the Radar is made.

Track 1: the ten papers

• Window: arXiv submissions from 1 to 15 November 2026.
• Pool: 665 candidates from AI, machine learning, vision, language, HCI, signal processing, statistics, medical-physics and a few related categories, plus Hugging Face daily papers. 194 were in scope after a keyword relevance filter.
• Criterion: Hugging Face daily-paper upvotes, highest first. Papers not featured on Hugging Face rank after those, by keyword relevance, then recency.
• Split: of the ten above, six were ranked by upvotes and four by the keyword fallback. The six had between 1 and 4 upvotes, so treat their order as a weak signal. The four fallback papers are not claimed to be popular. Upvote counts are as of 15 November 2026, 09:28 UTC.
• Skipped on reading: one candidate, ranked first with 10 upvotes, was dropped because it audits biomedical research agents (a scientific-claims benchmark and a drug-safety assay) rather than digital health or clinical care. The next paper in the shortlist took its place, so the numbers above are shortlist ranks 2 to 11.

Track 2: the personal health records spotlight

• Papers: a second shortlist run over a longer window, 17 August to 15 November 2026 (90 days), because few preprints in a fortnight mention personal health records or health apps. The pool was 2,325 candidates (arXiv results plus Hugging Face daily papers, before the scope filter). 9 were in scope: a paper needed a PHR or health-app term such as “personal health record”, “patient portal” or “mobile health app”, plus general health terms. Papers already in Track 1 were excluded. None of the nine was featured on Hugging Face daily papers, so all were ranked by keyword relevance and the top three are shown. Run as of 15 November 2026, 10:02 UTC.
• Apps: not a popularity or download ranking, because we cannot verify app-store data. The rule is: one PHR or health-record app from each of India, the Middle East and the US, each with a recent update or announcement on the operator’s or government’s own pages in 2026, described only from those pages. This issue picked one app each from India (Aarogya Setu 2.0), the Middle East (Sehhaty, Saudi Arabia) and the US (Google Health app, whose medical records sync Google describes for the US), the three markets we are tracking in the near term. A featured app is not repeated in the next three Radar issues unless it has a new announcement.

Checking

• An independent reviewer, a separate AI agent that sees only the draft and our written procedure and not our research notes, re-derives each factual claim from the primary source (arXiv for papers, the operator’s or government’s own page for apps), and checks links, upvote counts and the numbers in this section. It approves the text or lists fixes. There are up to three review rounds, and an issue that is not approved by then is not published unless the human editor authorises more.
• Papers are described from their abstracts and arXiv records, and results are reported as the authors’ own. Where an entry says so, details come from the full text: the three spotlight caveats, and for Bad Diagnosis also the dynamic analysis method and the permission and consent denominators. Anything labelled “our reading” is editorial opinion, not a finding. Apps are described as their operators describe them, and we have not tested them.
• An AI reviewer can miss errors.

---

The Research Radar is an alternate-day briefing from HCITExperts, alternating with The Clinical Frontier. Papers are linked above to their arXiv pages, and apps to their operators’ or governments’ own pages. Next issue in a couple of days.