Healthcare Technology AI/ML/DL

The Research Radar, Issue 002: Ten Digital Health and Clinical AI Preprints from 3 to 17 November 2026, a Personal Health Records Spotlight and Biomedical Research Agents

The Research Radar · 17 November 2026 · Issue 002 An alternate-day briefing on the ten highest-ranked recent preprints in digital health and clinical AI, plus a personal health records spotlight and a new section on biomedical research agents, written for India’s health-IT community. Preprints are early work. We say what the authors report, and where it stops.

In this issue

  • InclusionMed, the one paper this issue ranked by Hugging Face upvotes, is an open evaluation initiative. Its authors report index scores from 17.5 to 49.1 out of 100 and, per the full text, say the first release evaluates 22 models and has no benchmark for patients and the public.
  • Several of the ten papers are evaluations or audits of models or benchmarks, for example InclusionMed (a benchmark index), an audit of mental health recommendations from 41 language models, and a leakage audit of a chest X-ray benchmark.
  • A personal health records spotlight: three preprints, and three live PHR or health-record apps, one each from India (Eka Care), the Middle East (Shifa, Oman) and the US (Apple Health), described from their operators’ own pages.
  • New: biomedical research agents. Three preprints on agents used for biomedical science, a review, an automated chemical synthesis system and a computational benchmark. They are research tools, not clinical care.

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, 1 to 10. No candidate was skipped on reading.

1. InclusionMed: Benchmarking AI for Inclusive Healthcare

The InclusionMed Collaboration · arXiv:2611.21218 · 16 November 2026 · Rank basis: Hugging Face upvotes (58, as of 17 November 2026, 15:23 UTC)

  • What it does: an open medical AI evaluation initiative. Its atlas organises benchmarks across three healthcare roles and three AI capabilities. From 388 candidate benchmarks the authors selected four public benchmarks, and added a private diagnostic reasoning benchmark of their own, then aggregated results for frontier language models into an index.
  • What it reports: overall index scores range from 17.5 to 49.1 out of 100 (full text: across 22 models from 11 families), and rankings differ by role and capability. On one benchmark (HealthBench Professional, full text), all ten top-ranked models scored higher on patient communication than on clinical decision-making.
  • Caveat: the abstract says the five benchmarks cover three of the atlas’s nine cells. Per the full text, there is no benchmark for patients and the public, the private benchmark is graded by a language model, and the authors say suitability for a clinical deployment still needs validation in the intended setting. The PDF header shows it is a technical report, version 0.5.
  • Why it could matter in India (our reading): an index like this can help a buying team see which role and capability a benchmark actually tests before they rely on a headline score.

2. Evidence-Conditioned Completion and Synthesis of Electronic Health Records

Sicheng Shu, Parker Knight, Huitao Li et al. · arXiv:2611.19546 · 14 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: EHRDiB, a diffusion-bridge model that either completes a partly observed EHR visit or synthesises a new one, while varying how much of a patient’s real information is supplied as input, and measures both downstream usefulness and disclosure risk.
  • What it reports: on MIMIC-IV and Duke Health data, the authors report that for synthesis the gap in a prediction metric (AUPRC) against real data is 45.2% smaller on ten MIMIC-IV targets and 53.3% smaller on numerical Duke data than the best comparator, and that a membership-inference attack score rose from 0.516 to 0.598 as more patient evidence was supplied.
  • Caveat: per the full text, the Duke data are not shareable, the authors say the method offers no formal privacy guarantee, results have no confidence intervals, the work covers single visits only, and code is promised on publication. Preprint.
  • Why it could matter in India (our reading): the privacy result is a reminder to test synthetic records for leakage before anyone treats them as safe to share.

3. CONTEXT-PCa: Evaluating the role of clinical context in MRI-based AI for detecting clinically significant prostate cancer (study protocol)

Shiqi Huang, Weixi Yi, Shaheer Ullah Saeed et al. · arXiv:2611.19082 · 14 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: a study protocol. The authors plan to test whether adding clinical variables to a published image-only AI model, with its weights fixed, improves detection of clinically significant prostate cancer on bi-parametric MRI. The abstract describes a cohort of 6,022 cases from 4,540 patients at University College London Hospitals NHS Foundation Trust, 43 clinical variables, a primary endpoint of the difference in AUROC, and five external cohorts. The full text calls it a retrospective cohort study.
  • What it reports: no results. It describes the planned design.
  • Caveat: a protocol, so there are no findings to cite. Per the full text, the authors list limitations such as a single tertiary centre, verification bias, labels that include presumed-benign cases, and that patient overlap with the image model’s development cohorts cannot be excluded.
  • Why it could matter in India (our reading): a published protocol is a template for how a hospital could pre-register an evaluation of a clinical AI tool before running it.

4. Multimodal Fusion: A Review of Breast Cancer Advanced and Traditional Modalities

Jawaher Al-Tuweity, Joao Paulo Papa · arXiv:2611.18793 · 13 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: a review of imaging modalities used in breast cancer detection and of AI-based multimodal fusion.
  • What it reports: per the full text, the authors searched a literature database for 2017 to 2025, identified 1,050 papers plus 2 from citation search, and included the 50 their ranking placed highest. They state gaps in the literature: small private or single-centre datasets, limited prospective multicentre evidence and limited interpretability.
  • Caveat: a review. Its statements describe the literature it reviewed, not new evidence. The arXiv record gives no peer-review status.
  • Why it could matter in India (our reading): the gaps the authors list are a checklist of questions to ask a vendor of a multimodal breast imaging tool.

5. Understanding Users’ Security, Privacy, and Trust Concerns in Healthcare AI Systems

Tushar Nayan, Mohammed Al Duniawi, Song Fang et al. · arXiv:2611.20223 · 15 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: an analysis of discussions in 14 medical Reddit communities to see what security, privacy and trust concerns people express about healthcare AI.
  • What it reports: the abstract cites 6,514 relevant posts, which the full text explains are threads a classifier labelled as security, privacy and trust related. Per the full text, 46.02% of the labelled discussion fell in a “Trust and Responsible AI” theme, and around the events the authors timed, 5 of 12 AI development events, 2 of 5 privacy incidents and 4 of 5 regulatory events showed a statistically significant change in discussion.
  • Caveat: per the full text, Reddit users are self-selected and pseudonymous with unverifiable roles, the data are English only, much labelling is automated, and the authors describe changes in discussion as coinciding with events. The arXiv record gives no peer-review status.
  • Why it could matter in India (our reading): it samples English-language online talk, so it is a prompt for a local survey of patient trust, not a substitute for one.

6. From AI-Enabled Remote Patient Monitoring Toward Agentic Robotics: Embodiment, Governance, and Care Integration

Ivana Ragusa, Lorenzo Guiducci, Filippo Cavallo · arXiv:2611.18484 · 13 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: a scoping review of AI-enabled remote patient monitoring systems, which then proposes “agentic robotics” as a supervised, auditable and interruptible direction.
  • What it reports: per the full text, the review covers 93 core system papers drawn from 516 records. Of these, 4 (4.3%) reached sustained real-world deployment, 22 (23.7%) reached the two higher levels of decision-action coupling, and 16 (17.2%) were embodied systems.
  • Caveat: the abstract proposes agentic robotics as a future direction, and the paper is a review that reports no test of it. Per the full text, the authors say the embodiment finding applies only to the monitoring corpus and should not be generalised to healthcare robotics as a whole. The arXiv comment says it is accepted for publication in Annual Review of Control, Robotics, and Autonomous Systems, Volume 10 (2027).
  • Why it could matter in India (our reading): the deployment count is a useful sanity check on vendor claims about autonomous home monitoring.

7. Curbside Chatbots: Auditing the Mental Health Interventions Large Language Models Recommend to Clinicians, Caregivers, and Patients

Matthew Flathers, Jill Noorily, Phuong Anh Nguyen et al. · arXiv:2611.12814 · 9 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: an audit of what 41 language models recommend for mental health conditions, using vignettes written in three voices (clinician, caregiver and patient).
  • What it reports: the abstract gives 19 vignettes and 35,055 responses. Per the full text, the vignettes are 17 conditions and 2 controls, and a language-model judge extracted 494,328 interventions from the responses.
  • Caveat: per the full text, the vignettes were written by the authors without patient data, there is no answer key, the judge model is itself one of the audited models, queries went through US accounts in American English, and the voice effect is confounded with whether the diagnosis was named. The PDF header says the paper is a preprint not yet peer reviewed, and the dataset is promised on publication.
  • Why it could matter in India (our reading): it uses English-language prompts sent from US accounts, so a clinician-led check in Indian languages is the open question for any mental health assistant here.

8. Target Recoverability under Selective Labels: A Failure Mode in ECG-Based Acute Coronary Syndrome Prediction

Alexander Schubert, Anton Hakansson, Ziad Obermeyer · arXiv:2611.11063 · 8 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: a study of how training only on tested patients can teach a model associations that fail or reverse in untested patients when a diagnosis target includes cases the input cannot show. The authors illustrate it in simulation and in ECG data.
  • What it reports: per the abstract, among patients who were not tested, three ECG foundation models fine-tuned on revascularization scored AUROC 0.381 to 0.478 (below 0.5) for 30-day major adverse cardiovascular events, while models trained on ECG-visible subtypes scored 0.680 to 0.709 and a model trained on unstable angina scored 0.310 (below 0.5).
  • Caveat: per the full text, the authors call 30-day major adverse cardiovascular events an imperfect proxy for potentially missed acute coronary syndrome. The arXiv comment says it is accepted to the Proceedings of Machine Learning for Health (ML4H) 2026.
  • Why it could matter in India (our reading): any hospital training on “who got tested” labels could ask the same question of its own data.

9. Leakage- and Label-Aware Benchmarking of Class-Imbalance Strategies for Multimodal Chest X-Ray and EHR Classification

Dang-Khoa Bui, Quang-Huy Tran, Minh-Khoi Nguyen-Bui et al. · arXiv:2611.07950 · 5 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: re-audits an earlier benchmark on MIMIC-CXR and MIMIC-IV, shows that splitting at record level leaked patients across splits, then re-benchmarks with patient-separated splits and refined labels.
  • What it reports: the abstract says about a third of radiographs carry labels that differ from the codes of their own admission. The abstract reports that, under patient-separated splits and in the evaluated configurations, no class-imbalance strategy raised the observed mean macro F2 over no augmentation by more than 0.01.
  • Caveat: per the full text, the refined labels are not clinically verified, the study uses one cohort from one hospital system, and the authors say the operating points do not establish screening value. Preprint.
  • Why it could matter in India (our reading): it is a short checklist for any team about to quote a validation score from a public imaging benchmark.

10. How Deep is the Signal? Time-Context up to 30 Years for Colorectal Cancer Prediction in a Primary-Care EHR Model

Achille Raviez, Lan Wang, Brendan Delaney · arXiv:2611.07265 · 5 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: tests how far back in a patient’s history a primary-care EHR foundation model should look when predicting colorectal cancer, from 2 to 30 years, against a gradient-boosting baseline. The abstract names North West London primary care.
  • What it reports: the abstract gives an AUROC of 0.551 at 2 years, 0.590 at 18 years and a baseline peak of 0.589. Per the full text, the transformer’s AUROC was 0.580 at 30 years and the baseline peaked at 14 years.
  • Caveat: per the full text, cases and controls are matched on age and sex, so the authors say the figures are not comparable with clinical tools such as QCancer, the authors list a single-region study, and the data are not public. The arXiv comment says it is accepted at ML4H 2026, Findings Track.
  • Why it could matter in India (our reading): this prompts asking how much history a model needs before paying to digitise decades of paper records.

Patterns across the ten

  • Several of the ten are evaluations or audits of models or benchmarks, for example InclusionMed (a benchmark index), Curbside Chatbots (an audit of 41 language models) and the leakage audit of a chest X-ray benchmark.
  • Three of the ten are not new experiments: a review (breast cancer fusion), a scoping review (remote patient monitoring) and a study protocol (CONTEXT-PCa).
  • InclusionMed’s authors report that of the 338 benchmark tasks where they could identify a geographic setting, 318 were associated with the United States, and that 88% of the 388 candidate benchmarks included English (full text).

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. Only one of the three preprints is about a PHR as such. The other two are about patient-facing AI. The apps are described from their operators’ own pages, and we have not tested them.

Three preprints on PHRs and patient-facing health AI

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.

Medical Fine-Tuning Reshapes, Not Reduces, Patient-Facing AI Risk: A Rubric-Based, LLM-Judge Evaluation Framework

Savan Doshi · arXiv:2610.24807 · 17 October 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: proposes a rubric that scores language model answers to patient questions by risk-bearing language, such as treatment directives, dosage expressions and overconfident assertions, rather than by factual correctness, and applies it to six open models, three general and three medically adapted.
  • What it reports: the abstract describes a rubric of six risk categories applied to six open models (three general, three medically adapted) on 200 real patient questions from the K-QA benchmark, and says two of the six categories had low agreement among the language-model judges. Per the full text, the six models form three medical-general pairs, mean overall risk scores differ negligibly across them (within 1.24 to 1.43 on a 0 to 5 scale, with overlapping 95% confidence intervals), and medical fine-tuning changes the profile of risk by category.
  • Caveat: per the full text, severity is scored by language-model judges, clinician calibration is described as ongoing, and the factuality comparison uses only 50 of the 200 questions. The arXiv record gives no peer-review status; the PDF header says it was presented at the DAIH Workshop at COLM 2026.
  • Why it could matter in India (our reading): it is an example of testing the kind of language a patient-facing assistant uses, not only whether its facts are right.

CG-Probes: Recovering Guardrail Directions from Patient Query Embeddings

Marko Rehacek, Vitezslav Dusek, Martin Rusinko et al. · arXiv:2609.31062 · 25 September 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: trains small probes on frozen text embeddings of Czech oncology search queries to score risk on three axes in patient queries (medical urgency, psychological urgency and topic sensitivity), and compares them with language model graders.
  • What it reports: the abstract describes probes competitive with open-weight language models on urgency axes. Per the full text, one axis was not recovered by a single direction, the probes missed none of 28 silver-labelled emergencies but escalated 19.5% of 133 benign queries, and only the frontier model separated clearly, on psychological urgency.
  • Caveat: per the full text, about 65% of the highest-grade medical urgency evaluation queries and 97% of the highest-grade psychological urgency queries are synthetic, the two annotating oncologists are co-authors, and the probes are not deployed. The arXiv comment says it is accepted as a short paper at CIKM ‘26.
  • Why it could matter in India (our reading): guardrails for patient queries will need local languages and local escalation paths, which this Czech study does not test.

Review Before Trust: Source-Grounded Integrity Gates for AI-Assisted Personal Health Records

Nora Girda, Adrian Groza · arXiv:2608.29965 · 30 August 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: a gate for a personal health record application in which values extracted from medical documents by a language model stay provisional until a deterministic monitor checks them against the source text. The authors say it evaluates system integrity rather than clinical correctness or clinical safety.
  • What it reports: the abstract reports a replay on nine laboratory reports with 102 manually labelled rows, in which a schema-only policy admitted 97 numeric candidates, while the stricter quote-and-row gate admitted 72 and sent 25 to review.
  • Caveat: per the full text, the data are nine reports from one adult member of the research team, labels were not clinician-adjudicated, extraction used a hosted model with no zero-retention claim, and a refusal does not show a value is false. The arXiv record gives no peer-review status; the PDF carries a PSCare 2026 proceedings header.
  • Why it could matter in India (our reading): any ABDM-linked app that digitises lab reports with AI could ask where a value is held until someone has checked it.

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, described from its operator’s own pages. The selection rule asks for a 2026 update or announcement on those pages. Apple’s pages meet it. The Oman Ministry page carries only a footer stamp, “Last updated: Mar 12, 2026”, which does not say what changed, and Eka’s 2026 post mentions its PHR only in passing. We say so in each entry.

Eka PHR: a private company’s records app, India

Sources: Eka Care’s post of 27 April 2026 and its medical records features page, which shows no content date (its footer reads “© 2026 Eka Care”).

  • What it is: the 27 April 2026 post, written by Eka’s AI team, is mainly about the company evaluating an NVIDIA model for agentic, multimodal workflows. In passing it describes Eka PHR as “a personal health record used by millions of Indians to consolidate prescriptions, labs, vaccinations, and discharge summaries into a single, portable, ABDM-compliant longitudinal record.” That is Eka’s own description.
  • Capabilities, as Eka’s page describes them: bringing health records from different places together, turning them into digital files automatically, connecting with Apple Health and Health Connect, a “Secret Locker” for documents that the page says are stored only on the user’s device, family records and an insurance policy view.
  • Shipped or planned: the page describes features in the present tense, and because it shows no content date we cannot say when any feature shipped. The page calls the product the Eka Care app. The 2026 post names no planned PHR features. Its stated plans concern agentic clinical workflows and a vernacular ambient scribe.
  • Scale: the 2026 post says “millions” and gives no count.
  • Our reading: this is a company’s app that the company says is ABDM-compliant. We have not verified that claim from a government page.

Shifa app: a Ministry of Health app, Oman

Sources: the Ministry’s Shifa App page, whose footer reads “Last updated: Mar 12, 2026”.

  • What it is: the page describes the Shifa App as “the community gateway for the National Unified Health Record project, serving as a central digital platform across all Ministry of Health institutions.”
  • Capabilities, as the page lists them: personal information, medical reports, medical history, a dependents profile, medications, laboratory tests, immunizations, vital signs, appointments, referrals between institutions, organ and blood donation, virtual consultations and chats, self-registration for visits with electronic payment, multiple profiles, and protection of the health record with a dedicated password.
  • Shipped or planned: the page has an “Upcoming Updates” heading. Planned items include a redesigned interface, a comprehensive health report request, blood donation booking and in-app support requests, and expanding electronic integration to include public and private healthcare institutions in a phased plan towards full interoperability within the Unified Health Record system.
  • Scale: the page gives no app usage figures.
  • Our reading: the footer stamp is page-level and does not say what changed or when any feature shipped, so treat the list as what the Ministry describes.

Apple Health app: a consumer platform, United States

Sources: Apple’s release of 9 September 2026, its release of 3 November 2026, and the CMS page for Apple’s pledge.

  • What it is: Apple’s 9 September release says “Health app users can already import existing labs and see their results and trends over time.” It also describes a redesigned Health app “available later this year, starting in U.S. English, with more languages to follow.”
  • What has shipped, and what has not: Apple’s 9 September release says users can already import labs. Per Apple’s 3 November release, the redesigned app is available in the iOS 27.3 developer beta, as a free update later this year. Apple’s September release describes a Longevity tab in the redesigned app, not yet released to the public as of 3 November, that draws on data from “iPhone, Apple Watch, AirPods, and clinical health records, as well as third-party apps and devices.”
  • Languages: Apple’s 3 November release lists global English, Chinese (Simplified and Traditional), French, Japanese, Korean, Norwegian, Portuguese, Spanish, Turkish and Vietnamese for the redesigned app.
  • Records access plans: the CMS page says “These capabilities are currently under development and expected in the future,” in reference to record access and sharing enhancements, with no launch date.
  • A limit on what we confirmed: we could not read a primary page that describes Apple’s separate Health Records feature, so we describe only what the pages above say.
  • Our reading: the records features sit inside a larger health and fitness app, and the redesign was in a developer beta as of Apple’s 3 November release.

Across the three apps and three preprints (our reading)

  • Each app’s source page describes health records alongside other functions: Eka lists family records and an insurance view, Shifa lists appointments and virtual consultations, and Apple’s release describes a health and fitness app.
  • Two of the three sources describe connections to other apps and devices: Eka’s page lists Apple Health and Health Connect, and Apple’s release refers to third-party apps and devices.
  • The three preprints address different questions: the risk language of answers to patient questions, guardrail risk axes for patient queries, and a gate that holds model-extracted lab values until checked against source text.

Biomedical research agents

New from this issue. These are preprints on AI agents and foundation models used in biomedical science, such as drug design and chemistry. They are research tools, not clinical care. We say what each paper does, what the authors report, and what kind of evaluation it was, from the full text. Three preprints came from a separate shortlist (see “How this issue was made”). None was featured on Hugging Face, so all were ranked by keyword relevance.

1. AI for Biologists: The Knowledge Boundaries of Foundation Models in the Age of Agentic Systems

Tian Cai, Andrew Kleist, Jaimin Patel et al. · arXiv:2611.17015 · 12 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: a review. It explains how biological foundation models are built, proposes five design elements that it says define a model’s “knowledge boundary”, and proposes standardised descriptions of model limits.
  • What it reports: no results of its own. Per the full text, it quotes other studies, for example one in which, on variant classification from DNA, a language model alone reached 49% and 80% with an added DNA model.
  • Caveat: evaluation type: review, with no experiments, simulation or wet lab. Per the full text, the paper proposes a “boundary statement” for each model and a community-maintained registry, and the acknowledgments disclose AI assistance in literature collection, fact verification and language editing. The arXiv comment says it is to be published in Annual Review of Biomedical Data Science.
  • Why it could matter in India (our reading): the idea of a model card for what a biology model cannot do is easy to borrow for any research group choosing tools.

2. Vibe synthesis: hierarchical agentic decision-making for intent-driven multistep molecule synthesis

Jie Chen, Guikun Chen, Feng Wan et al. · arXiv:2611.18772 · 12 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: VibeSyn, a hierarchical agent system that turns chemist-defined molecular objectives into decisions executed through computational tools and laboratory automation.
  • What it reports: per the full text, in a campaign reproducing a 10-step synthesis of catharanthine first reported in 1987, human input accounted for 7.3% and human confirmation for 4.2% of 722 recorded actions. For the cyclic peptide campaigns the figures were 7.5% and 5.2%.
  • Caveat: evaluation type: wet-lab demonstrations of chemical synthesis. Per the full text, there is no baseline comparison, humans submitted tasks, a human chemist ran one step by hand, and nothing was tested biologically or clinically. The arXiv comment says “30 pages, 5 figures”.
  • Why it could matter in India (our reading): it shows what “human in the loop” can mean in numbers for an automated laboratory.

3. Pyrite: Building and Benchmarking Protein-Ligand Pose Ranking Agents

Megha Srivastava, Michael H. Lim, Allison M. Keys et al. · arXiv:2611.14511 · 10 November 2026 · Rank basis: not featured on Hugging Face, ranked by keyword relevance

  • What it does: a benchmark of 1,500 pairs of computed ligand poses, and two language-model agents that rank candidate poses pairwise.
  • What it reports: the abstract reports agent accuracies of up to 67% and 74%, exceeding observed expert performance. Per the full text, on the 265-pair curated split with human annotations, five expert chemists reached 0.45 accuracy counting abstentions as wrong and 0.57 on decisive answers only, and a plain logistic regression scored 0.64.
  • Caveat: evaluation type: computational (in silico) benchmark against public crystal structures, with no patients and no wet lab. Per the full text, the candidate poses come from the authors’ proprietary model, the task is pairwise and the authors list transfer to real ranking pipelines as future work. The arXiv comment says it is a NeurIPS 2026 Agentic AI for Biological Discovery Workshop paper.
  • Why it could matter in India (our reading): a plain logistic regression scored 0.64 on the same curated split (the agents reach up to 67% and 74%), so asking for simple baselines before paying for an agent is a reasonable habit.

The takeaway

Read these as a sample of recent preprints chosen by a stated method, not as evidence of what works in a ward. Several of this issue’s papers are evaluations or audits of models or benchmarks: InclusionMed’s authors say the five benchmarks in their Atlas cover three of its nine cells and, per the full text, that it has no benchmark for patients, and the remote monitoring scoping review counts, per the full text, 4 of 93 systems at sustained real-world deployment. The new biomedical research agents section is about research tools, not clinical care. We will repeat this on alternate days with a fresh window.

How this issue was made

Process version 3, 17 November 2026. Version 1 put the method note at the top of the issue and had no personal health records section. Version 2 moved the method here and added the personal health records spotlight. Version 3 adds a separate section on biomedical research agents, with its own shortlist and rules. This section will record future changes to how the Radar is made.

Track 1: the ten papers

  • Window: arXiv submissions from 3 to 17 November 2026.
  • Pool: 791 candidates from AI, machine learning, vision, language, HCI, signal processing, statistics, medical-physics and a few related categories, plus Hugging Face daily papers. 233 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, one was ranked by upvotes (InclusionMed, 58) and nine by the keyword fallback. The nine are not claimed to be popular. Upvote counts are as of 17 November 2026, 15:23 UTC.
  • Skipped on reading: none.

Track 2: the personal health records spotlight

  • Papers: a second shortlist run over a longer window, 19 August to 17 November 2026 (90 days), because few preprints in a fortnight mention personal health records or health apps. The pool was 2,341 candidates. 7 were in scope: a paper needed a PHR or health-app term plus general health terms. Papers already in Track 1 were excluded. None of the seven was featured on Hugging Face, so all were ranked by keyword relevance and the top three are shown. As of 17 November 2026, 15:24 UTC. Only one of the three is about a PHR as such. The other two are about patient-facing AI.
  • 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 Eka PHR (India), the Shifa app (Oman) and the Apple Health app (US). The Oman Ministry page carries a footer stamp of “Last updated: Mar 12, 2026” but does not say what changed, and Eka’s 2026 post mentions its PHR only in passing, so for those two entries the update-or-announcement test is met only loosely. We kept them because their primary pages were readable. Apps featured in Issue 001 (Aarogya Setu 2.0, Sehhaty and the Google Health app) are not repeated unless they have a new announcement.

Track 3: biomedical research agents (new)

  • Papers: a third shortlist over 3 to 17 November 2026 (14 days). The pool was 320 candidates, and 21 were in scope. A paper had to match both an agent term and a biomedical-science term, so purely clinical papers and generic agent papers do not qualify. The Track 1 and Track 2 shortlists were excluded. None of the 21 was featured on Hugging Face, so all were ranked by keyword relevance and the top three are shown. As of 17 November 2026, 15:25 UTC. No paper was brought back from an earlier skip.
  • Framing: these are research tools for biomedical science, not clinical care. The first of the three is a review, and the second is chemistry, so the section’s scope is broad.

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 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 “per the full text”, the detail comes from the full text: the cohort, data and limitation details marked in the entries. 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’ own pages. Next issue in a couple of days.

Share: Twitter LinkedIn