Changelog

KM screen settings now persist in the browser: the X-axis title per time column, the Y-axis title per event column, and the look (font, at-risk table, confidence bands, P-value position) per user. They are reused the next time you open the screen and when the submission batch regenerates the figure.

Figure visual QA can now apply its findings with one click:

• KM axis titles (missing unit, raw column name, wrong endpoint name): saved per column; if the recipe is confirmed, the KM curve is regenerated automatically
• Sankey stage labels (raw codes 0/1/2): saved to the group-label dictionary; the chart is reopened with the same settings and saved
• KM X-axis maximum (too much empty space on the right): saved per time column and regenerated

Also: the manual Categorical / Continuous choice made in Cox or Table 1 is stored in the recipe and shared by both screens, and integer-coded columns whose name says they are a category (subtype, stage, grade…) are treated as categorical even with many levels (previously 7+ levels were modelled as a continuous "HR per unit").

Two buttons were added to the progress board. Statistical review reads your recipe (column roles), the automatic checks (events per variable, proportional-hazards test, VIF, events per group, missingness) and the numbers of the saved analyses (Table 1, KM, Cox, forest plot, propensity-score balance) as the target journal's statistical reviewer would, and returns a Japanese summary with English reviewer comments, a concrete fix and the evidence for each comment, suggested sensitivity analyses (competing risks, a different covariate set, OS instead of RFS, PSM/IPTW, E-value, cut-off exploration) and a draft 'Sensitivity analyses' paragraph for Methods. Each suggestion has a Run button: the recipe is changed for that run only, the Table 1 / KM / Cox batch runs, and the recipe is restored afterwards (results are saved as sens_… and never mixed with the main analysis). Figure QA shows the latest figure of each kind (KM, Figure Composer, forest, flow diagram, chart) to the AI as an image and compares it with the numbers in its own .rko: P values, n and risk-table counts printed on the figure, axis titles and units, legend labels vs group labels and raw codes, clipped or overlapping text, pixel size against the journal's dpi requirement, and consistency across figures. Neither sends row data (review 40–80 credits; figure QA 15–30 credits per figure).

In the data-profile screen that opens after loading a file, type one sentence (e.g. 'Compare recurrence-free survival by pleural invasion, adjusted for age, sex and stage; target journal JTCVS') and the AI proposes column roles (time, event, competing, group, covariates, excluded), the journal, study design, primary and secondary endpoints, the planned outputs and English group labels, and applies them (it avoids mostly-missing derived columns, prefers months over days, and keeps covariates to the 5–20 that belong in Table 1). Nothing else needs typing: clarifying questions come with answer chips, the role summary has dropdowns for time / event / group and chips with × and '+ add' for covariates, and a free-text box ('exclude stage IV') asks for a revision. When a multi-level column such as pathological stage IA1–IIIB must be compared as 'I vs II/III', the AI proposes a derived group column with its mapping; the mapping is edited by clicking (unmapped values become missing), and on confirm the column is added to the data, used as the group and included in exports. Confirming then runs the submission set (Table 1, KM, Cox) in one go. Only column names, types, levels, missingness and legends are sent — no row data and no columns that look like identifiers (5–15 credits per run).

Where a group has not yet had an event, the 95% CI used to print as "100.0% (100.0–100.0%)" (0.0–0.0% in cumulative mode): the Greenwood variance is zero there, and a log-log transform does not help. Such time points now use the two-sided 95% Clopper-Pearson exact interval based on the number at risk at that time (lower limit = 0.025^(1/n) ≈ 1 − 3.7/n). We verified in R that this equals what the beta product confidence procedure for censored data (Fay, Brittain & Proschan, Biostatistics 2013) returns at zero-event times. The familiar "rule of three" (3/n) is one-sided 95%, so we chose the two-sided version to match the other CIs in the figure. The correction applies to the CI band, the time-point annotation and the time-point table; affected values carry a dagger and a footnote is added to the figure when a group has no events. The same rule fills in the upper limit for cumulative incidence (CIF) when the event of interest has not occurred. A display option restores the previous behaviour. The .rko records the CI method (ciMethod) and the dagger flags, so AI-drafted Methods describe it correctly.

Three chart types were added. Sankey diagrams show patient flow across ordered stages such as stage or treatment line, with node names editable inline. Waterfall plots show each patient's best change from baseline (e.g. RECIST). The two-arm outcome comparison draws group A vs B proportions with 95% CIs and a P value, imports from KM / CIF .rko files, and has a style panel with journal-style axis ticks, one-decimal percentages and P formatting. Visual Abstract gains two layouts for The Annals of Thoracic Surgery (a1 / a2) with an Annals theme: cohort and methods on the left, the figure in the middle, key results on the right, an A-vs-B bar for secondary outcomes, and aspect-preserving figure export. A Sankey .rko loaded as the figure passes its flow counts into the AI digest. Also fixed: multi-panel Figure Composer .rko files keep all panels, and typing an existing account into the sign-up form no longer dead-ends.

195 pictograms were added to the free library, for 282 in total. Beyond study design (retrospective / prospective, crossover, blinding, placebo, ethics approval, consent, follow-up, screening funnel, timeline, endpoint, animal model, in vitro), diagnostics (chest X-ray, MRI, PET, ultrasound, ECG, spirometry, biopsy, pathology, glucose, urinalysis, mammography, retina, culture, sequencing, biomarker), treatment (targeted therapy, immune cell, transfusion, anesthesia, rehabilitation, palliative care, vaccine, insulin, eye drops, cast, wound dressing, counseling), people (nurse, pharmacist, family, pregnancy, fetus, newborn, home care, ambulance, discharge, team, older adults, community) and charts (scatter / regression, ROC, box plot, pie, table, heatmap, error bars, confidence interval, trends, flowchart, check / cross, arrows, star), the set now covers cardiology (heart rate, arrhythmia, valve, aorta, AED, CPR, cardiac surgery), gastroenterology (appendix, stoma, hernia, cirrhosis, gastric ulcer, colonoscopy, spleen, gallbladder, small intestine), neurology and psychiatry (stroke, seizure, dementia, neuron, headache, tremor, mental health, stress, mindfulness), orthopedics (fracture, hip, disc, osteoporosis, crutches, walker, joint), obstetrics and gynecology (ovary, menopause, IVF, infant care), urology (stone), dermatology (rash, skin lesion, burn), ophthalmology and ENT (glasses, cataract, nose, throat, hearing), endocrinology and hematology (hormone, scale, blood cells, bone marrow, coagulation), infectious disease (hand hygiene, vector, antimicrobial, isolation), general oncology (metastasis, response, cell division, mutation, chromosome, gene editing, staging), emergency care (intubation, trauma, vital signs, first aid) and thoracic surgery (lobectomy, segmentectomy, pneumonectomy, lymph node dissection, pleural effusion, pneumothorax, lung nodule, mediastinum, esophagectomy, tracheostomy). The library page now filters by category (people / organs / disease / procedures / treatment / diagnostics / devices / study design / charts / lifestyle). The same 282 are available to the Visual Abstract AI picker and icon list. Still free, unlimited and sign-in free.

New Subgroup and Charts pages, and a more rigorous treatment of competing risks and weighted analyses. Subgroup re-fits the treatment effect within each stratum (continuous variables split at the median) under Cox, Fine-Gray or logistic models and draws one forest plot (P for interaction by likelihood-ratio test, off by default). Charts gives grouped, stacked, 100 % stacked and year-by-year trend plots with chi-square / Fisher chosen automatically. Both inherit the adjusted cohort and can be dropped into Figure Composer panels. Fine-Gray now takes IPTW / overlap weights (robust variance) and can add bootstrap 95 % CIs for the sHRs (with progress and stop). A 0/1/2 competing-event column is refused under plain KM or Cox with guidance, and switching to CIF / Fine-Gray swaps the event column automatically. Also: Japanese labels in KM figures, Table 1 continuous variables as mean (SD) / median [IQR] / automatic by normality, E-values for unmeasured confounding in Cox, a copyable view of the R code that ran, and a component table for composite endpoints (MACE).

A set of features to cut the busywork: the data profile after import sets column roles and the target journal once, and KM, Cox, Table 1 and adjustment all use them (analysis plan, stored in the browser). "Build submission set" on the progress board runs Table 1, KM and Cox automatically and saves the .rko files; "Health" lists events, EPV, proportional hazards, VIF and matching rate. Propensity-score matching now runs on MatchIt (nearest / optimal / full, caliper, exact), and the matched cohort is shared by Table 1 and Cox (pair-clustered robust variance). Also: a variable dictionary, an analysis log, flow-diagram counts from data, a submission export preset, Core Draft batch generation and a ⌘K command palette.

87 clinical research pictograms are now available: line art of organs, procedures, treatments and study designs, exported as SVG or PNG in any colour (five palettes plus custom) and stroke width. Free, unlimited, no sign-in — with copy-as-image for PowerPoint and bulk ZIP download. Open it from "Pictograms" on the home page or the app toolbar.

A broad usability pass. Exports are now uniform across every page (PNG 300 DPI free; 600 DPI, TIFF, PDF and SVG on Pro), figure and Excel text stays English even on the Japanese UI, Cox / Table 1 column widths and forest-plot markers were tuned, and a crash on the diagnostics page for lower-is-positive markers was fixed. Remaining untranslated Japanese UI strings were completed and light-theme contrast was rechecked.

Visual Abstract gains journal-submission templates (four layouts: two-arm comparison, three-column, quadrant, and a portrait layout for social media). Paste the title and abstract and the AI drafts the content, copying numbers only from the abstract or your .rko results. A pictogram is picked for each arm automatically, and PowerPoint exports are editable shapes. Journal-specified layouts are added on request.

Every analysis page now has "✨ Set up with AI": describe the analysis and the time, event, group and confounder columns are assigned for you (KM, Cox, Table 1, change analysis, propensity adjustment). A free local auto-detect handles the common case; the AI steps in only for complex requests.

Double-click any text in your figures to edit it in place. Flow diagrams: every box, title and footnote updates live as you type. KM curves: rewrite group and axis labels right on the plot and redraw in one click — fixing raw "0/1" legends takes seconds. The KM grouping column can now be changed on the page too.

New: AI English labels for Table 1 (Pro). Japanese variable names and category codes (e.g. 性別 １：男性 ２：女性) convert to publication-style English (Sex / Male / Female) in one click, applied to every export (Word, Excel, .rko) — and you can switch back anytime. Complete Draft now picks the manuscript language automatically from the target journal's guidelines.

The diagnostic-accuracy suite (ROC, marker comparison, kappa, calibration, Bland-Altman/ICC) now exports publication-ready copies, .rko files and SVG. Detection-limit values (e.g. "<2.0") are now handled consistently across all pages and disclosed automatically in the Table 1 footnote.

Added the AI assist: type what you want to do and it suggests the right feature, opening the screen with columns pre-assigned (catalog search is free; AI suggestions are Pro). Also added design checks (in-browser detection of immortal time bias and more), automatic detection and exclusion of identifier columns, a full "what is sent to AI" preview, and a data-handling page with an ethics-review template.

Table 1 is now study-design aware: RCT and matched designs show an SMD column instead of p-values (with a CONSORT 2010-compliant footnote), and a missing-data column can be added. Cox regression gained model diagnostics (C-index, AIC, EPV warning, proportional-hazards test, overall p for factors) and a pre-specified covariate mode.

Major Adjust expansion: 1:k and with-replacement matching, weight trimming, PS-quintile stratification, PS-as-covariate, variance ratios and PS overlap, weighted logistic (OR), and sensitivity analyses (trimming sensitivity, Rosenbaum bounds, negative controls).

Added Adjust (confounding adjustment): propensity-score IPTW, ATT, overlap weights and 1:1 matching, balance diagnostics (love plot, adjusted Table 1), adjusted KM / CIF curves and E-values — all in in-browser R.

Added Complete Draft, with submission guidelines for 39 journals preset.

Added the STROBE check (beta). The AI reviews whether your manuscript follows the guideline.

Added Flow Chart. Build diagrams that follow CONSORT and STROBE.

RINKO.ai is now generally available. Statistical analysis through manuscript drafting in a single tool.