FAQ
Frequently Asked Questions
What is SpineCRS?
SpineCRS (Spine Complication Risk Score) is a clinical decision-support platform combining a preoperative wound-complication risk calculator with a standardised classification system for perioperative complications in spine surgery. It is developed at Balgrist University Hospital, University of Zurich.
Is the risk calculator validated?
The calculator is derived from a retrospective cohort of 4,067 consecutive lumbar spine surgeries (2017–2023) at a single academic centre. Seven independent predictors of major wound complications were identified through multivariate logistic regression. External validation is ongoing as the prospective registry accumulates cases.
How does the model improve over time?
SpineCRS uses a self-calibrating model pipeline. The published coefficients (Klossner et al., n=4,067) are stored as the seed version (v1) and load into the calculator dynamically, so the cohort estimate you see always reflects the current model. As the prospective registry grows, the model refits automatically: as new confirmed complication events accumulate, a fresh logistic regression is computed (iteratively reweighted least squares) over the combined retrospective and prospective cohorts, and its performance is evaluated — area under the ROC curve, Brier score, and decile calibration. The active version and the sample size behind it are always shown above the cohort estimate.
Does a recalibrated model go live automatically?
Only if it passes a set of promotion guard checks. Each candidate refit is compared against the active version on discrimination (AUC), calibration, and coefficient stability before it can be promoted; a candidate that degrades performance or drifts implausibly is held back rather than published. Every version is retained with its full metrics, so the model history can be reviewed and an earlier version restored at any time. The evolving model performance — version timeline, AUC and Brier history, coefficient drift, and calibration — is visualised on the Evolution page. As with all SpineCRS outputs, recalibrated estimates are provided for research, educational, and informational purposes only and are not intended to guide individual clinical decisions.
What is the institution network?
An optional network that lets surgeons at the same institution pool the de-identified characteristics of their cases, so the calculator can offer a risk estimate matched to that institution's case-mix rather than to the reference cohort alone. It is opt-in per surgeon: enable it under Profile → Institution network, and switch it off again whenever you like. Once the surgeons who have opted in at your institution have contributed 100 pooled cases between them, Individualize to my institution unlocks on the calculator. The Stats view shows how many colleagues are in your network, how many cases are pooled, and how many more are needed to unlock. Joining changes nothing about who can see your cases — row-level security still restricts every case row to the surgeon who entered it.
How is the institution-individualized risk estimate calculated?
Switch Individualize to my institution on and choose the covariates to match on — age, sex, ASA class, Charlson index, smoking status, procedure type, operating time, preoperative haemoglobin, and the seven model predictors. The server then fits a propensity model separating your institution's cases from the pooled reference cohort on exactly those covariates, keeps only the reference cases lying inside your institution's propensity support (the matched sub-cohort), and refits the same seven-predictor logistic regression on that matched set. The result appears beside the standard estimate with its own 95% confidence interval and the size of the matched baseline. Two safeguards apply. If the propensity model cannot meaningfully separate your institution from the reference cohort (c-statistic below 0.55), the cohorts are treated as equivalent, the whole cohort is used, and both estimates are identical — which is itself an informative result. If the matched sub-cohort falls below 300 cases or 30 complication events, no estimate is shown at all, because a refit on that little data would be unstable. Adding covariates tightens the match but shrinks the matched sample, so there is a deliberate trade-off between specificity and precision. Like every other SpineCRS output, the matched estimate is provided for research, educational, and informational purposes only; it is computed on demand and is not stored as a model version.
What exactly is shared with my institution network — can colleagues see my cases?
No. Only the de-identified clinical characteristics of cases entered by surgeons who have opted in are read, and only server-side, inside an authenticated function. Patient identifiers live in the separate owner-only table (case_local_refs) and are never touched by the network. Your browser receives only aggregate counts — how many surgeons have opted in, how many cases are pooled — and the coefficients of the matched model; never another surgeon's rows and nothing at case level. The aggregate summary comes from a SECURITY DEFINER function that emits counts only. Pooling is evaluated live from whoever is opted in at that moment, so opting out removes your cases from the pool immediately.
Can I use the calculator without registering?
Yes. The calculator is freely available for educational use without an account. Registration is required only if you wish to submit cases to the prospective research registry and track outcomes longitudinally.
How are patient data protected?
Patient identifiers (PIDs) are stored in a separate, owner-only table (case_local_refs) with row-level security restricting access to the registering surgeon. The main cases table contains only de-identified clinical data. No demographic or contact information is collected.
Where is the data hosted?
SpineCRS is operated on European infrastructure, with the clinical database and authentication services hosted in a Swiss data centre. The platform is operated in accordance with the revised Swiss Federal Act on Data Protection (revFADP) and the EU General Data Protection Regulation (GDPR). No clinical data leaves this infrastructure except for the limited, surgeon-initiated calls described under the auto-fill questions below.
What security measures protect my data?
All traffic between your browser and the server is encrypted in transit (TLS/HTTPS). The database is encrypted at rest. Access control is enforced server-side via row-level security (RLS): each surgeon can read and write only their own cases and patient identifiers — there is no shared workspace, and surgeons cannot see one another's data. Third-party API keys never reach the browser; they are stored as server-side secrets and only the de-identified payload is forwarded. Authentication uses standard JWT sessions with email verification, and sign-up is gated by institutional email-domain checks where applicable.
Can I fill Stage 1 automatically from chart screenshots?
Yes — Autofill from screenshots at the top of Stage 1 transcribes the form from chart images. Add as many as the case needs (preoperative sheet, anaesthesia record, operative report) by choosing files, pasting from the clipboard, dropping them in, or using your phone camera. Each image is then cropped on your device: you drag a box over the clinical panel, everything outside it is discarded before upload, and the re-encoding strips EXIF metadata (GPS, timestamp, device). Only the cropped regions leave your device, in a single call for the whole case, and none of them is stored. What comes back is a review in which every value is pre-filled and editable — no per-field confirmation, with an amber dot on anything the model was unsure about and an edited marker on values you changed. Two things block applying the result: patient age, which you always enter yourself and which is never inferred, and any field where two screenshots disagree — you must pick the correct value before continuing. The risk-cluster fields (BMI, CRP, blood loss, dural tear) are shown read-only, since those stay tied to the calculator snapshot for the case. Manual entry remains fully available; nothing about Stage 1 requires the extractor.
How does the automatic Charlson Comorbidity Index (CCI) scoring work?
The Charlson conditions are filled either by the Stage-1 screenshot autofill described above, or on their own: in Stage 1 you may fill the Charlson score manually or click Auto-fill. The auto-fill feature accepts a pasted diagnosis list or a photograph of one (for example, a printed problem list) and uses a clinical language model (Anthropic's Claude) to identify which of the 17 Charlson conditions are present. The result is shown as a checklist for you to verify and correct before applying it to the form. Patient age is entered separately and is never inferred by the AI — only you decide which age bracket applies. The final score and the 17 confirmed booleans are saved with the case for auditability.
What happens to the text or photo I send to the auto-fill?
Nothing is stored. This holds for both the Charlson auto-fill and the Stage-1 screenshot autofill. Pasted text and uploaded images (for screenshots, the cropped regions only) are forwarded over HTTPS to the language model for a single inference call and are then discarded. The server-side proxy never logs or persists the source material, and no column in the database holds the raw text or image bytes. Only the structured result you confirm — the 17 condition flags, the age points you entered, the chosen method, and a confidence tag — is written to the case record. The verbatim diagnosis list shown to you for verification exists only in the modal and is dropped when you close it.
Can I upload a photo with patient details on it?
No — please don't. Only upload photos that contain the diagnosis list itself and no patient identifiers. Crop, fold, or cover the patient name, date of birth, patient/case number, address, insurance information, and any institutional barcodes or stickers before taking the picture. The auto-fill feature only needs the list of conditions; everything else is unnecessary and should not leave your device. If in doubt, paste the diagnoses as text instead. When using Autofill from screenshots, the on-device crop step is there for exactly this purpose: place the box around the clinical panel only and leave the patient banner outside it — the discarded region never leaves your device.
Where is the AI processing performed?
The AI inference call goes to Anthropic's API endpoint (api.anthropic.com), which is operated on AWS infrastructure in the United States. This means that any text or image you submit to either auto-fill feature briefly leaves the EU/Swiss environment for the duration of the call. The institution network does not use AI and never leaves the Swiss/EU infrastructure — propensity matching and the refit run entirely in the database environment. The request and response are encrypted in transit, Anthropic does not retain API inputs for training, and — as noted above — nothing about the request is stored on SpineCRS servers afterwards. If your institutional data agreement requires that no clinical content leaves European infrastructure, please fill the Charlson score manually rather than using auto-fill.
Why is the blood-loss field empty when I'm logged in?
For prospective cases, the blood-loss field requires the actual measured intraoperative value rather than a population estimate. The automatic per-category estimate (Cat I ≈ 209 mL, Cat II ≈ 403 mL, Cat III ≈ 612 mL) is shown only in the open/educational calculator when no surgeon is signed in.
How do I join the registry?
Use the Create an account link on the sign-in screen. If your institution is already enrolled, sign-up is immediate (subject to email-domain verification). If not, the form lets you submit an enrolment request to the project coordinator.
What is the SpineCRS classification system?
SpineCRS classifies perioperative complications by four axes: surgical complexity (Cat I–III), cause (surgical vs. medical), therapeutic consequence (grades A–E), and neurological deficit. It was developed and validated against cumulative hospital stay in 934 consecutive spine surgery patients (Farshad et al., The Spine Journal, 2020).
Who can I contact?
For clinical enquiries, research collaboration, or technical issues, please contact the author via Balgrist University Hospital, Zurich, or by email at
contact@spinecrs.com. Institutional enrolment requests are routed through the in-app registration flow.
This FAQ will be updated as the platform evolves. Suggestions for additional questions are welcome.