Clinical Calculators: where the AI stops and the arithmetic begins | Cynaps Blog | Cynaps
Clinical Calculators: where the AI stops and the arithmetic begins
A library of thirty eight bedside scores across eighteen specialties - CHA2DS2-VASc, Wells, qSOFA, eGFR, MELD-Na, NIHSS, NEWS2, APGAR, PHQ-9, ECOG and more - computed by deterministic code that matches the published formula to the decimal, transparent down to each input, cited to the derivation paper, and wired to fire automatically when the clinical chat reaches a matching diagnosis.
Cynaps Editorial
June 9, 2026 · 5 min read
Clinical CalculatorsRisk ScoresClinical Tools
The score you half remember
Every clinician carries a mental shelf of bedside scores: CHA₂DS₂-VASc, Wells, qSOFA, CURB-65, the HEART score, GCS, eGFR. The trouble is rarely which one to use. It is the friction around it: recalling the exact point weights, opening another calculator site, retyping the patient's numbers into a form, and copying the result back into the note. A twenty second decision wrapped in two minutes of tab switching.
And in a product full of AI, there is a sharper worry: is a language model doing this arithmetic? For a stroke risk or transplant priority score, "probably right" is not good enough.
What it is
Clinical Calculators is a library of bedside scoring tools, each one a clean form: tick the boxes, enter the values, get a result. Thirty eight are wired into the gallery today, across eighteen specialties, grouped by where a clinician actually reaches for them:
Cardiology and VTE. CHA₂DS₂-VASc, HAS-BLED, HEART, Wells - PE and DVT -, PERC, RCRI, ASCVD PCE 2013.
Critical care and early warning. qSOFA, SOFA, NEWS2.
Pulmonology and infectious disease. CURB-65, Centor / McIsaac.
Gastroenterology and haematology. Glasgow-Blatchford, HScore.
Anaesthesia and perioperative. ASA Physical Status, STOP-BANG.
Endocrinology. Hyperglycaemia corrected sodium.
Bedside calculations and dosing. BMI, body surface area - Mosteller -, ideal body weight - Devine -, anion gap, albumin corrected calcium, maintenance IV fluids - Holliday-Segar -.
Mental health. PHQ-9, GAD-7.
Oncology. ECOG Performance Status.
Obstetrics and paediatrics. Bishop score, APGAR, Westley Croup.
Every result comes back the same way: the score, a risk tier, a one line interpretation, a recommendation drawn from the guideline where one applies, a component breakdown showing what each input contributed, and the reference to the study the score comes from - Lip 2010 for CHA₂DS₂-VASc, Six 2008 for HEART, Inker 2021 for the CKD-EPI 2021 equation, Lim 2003 for CURB-65, and so on -.
The same calculator definition drives both the form and the result page. Add a new score to the registry and the gallery, the form, and the result layout pick it up automatically; nothing has to be re rendered by hand. The form also accepts deep links with prefilled inputs, so a calculator embedded in a chat answer can land the clinician on the result with the patient's numbers already entered.
The point: the math is not the model's
This is the part that matters most, and it is deliberately boring: every calculation is deterministic, hardcoded arithmetic. No language model touches the number.
The CKD-EPI eGFR equation is the published 2021 formula, coded exactly: 142 × min(SCr/κ, 1)^α × max(SCr/κ, 1)^−1.200 × 0.9938^age × sex coefficient. CHA₂DS₂-VASc is the paper's point table - CHF +1, age ≥75 +2, prior stroke +2 -, summed in plain code. MELD-Na is its log based formula with the sodium correction and the UNOS sodium clamp at the boundaries. There is no temperature, no sampling, no "probably." The same inputs produce the same output every time, a result a clinician can reproduce on paper and audit back to the derivation study printed alongside it.
Inputs are validated before they are used. Each field has a documented range - age 18 to 120, creatinine 0.1 to 30 mg/dL, sodium 100 to 170 mmol/L, and so on -, and an out of range value is rejected with a clear, doctor readable message like "Field 'age' must be ≥ 18; got 10" instead of being silently coerced. A typo fails loudly rather than producing a confident wrong score.
A small but important nuance on ASCVD: the 2013 Pooled Cohort Equations are the published formula and the tool computes them exactly as written, including the binary race coefficient. The result page calls out that the 2023 PREVENT equations are now the preferred risk model where available, so a clinician using the older score sees the caveat alongside the number rather than reading them out of context.
Transparency, not a black box
A number without its workings is hard to trust at 3 a.m. Every result shows its component breakdown, the exact line items that produced the total, so a clinician can see why a CHA₂DS₂-VASc came out at 5 - CHF +1, hypertension +1, age 75 +2, prior stroke +2 -, not just that it did. The risk tier translates the number into a band - low, moderate, high -, and the recommendation, where present, is text lifted from the guideline itself - for example, "oral anticoagulation recommended" -, not a sentence a model improvised. The citation is right there to open.
They come to you
The calculators are not only a destination you navigate to. They are wired into the clinical chat.
When the engine forms a working diagnosis, the relevant score fires automatically, pulling what inputs it can from the case and surfacing the result alongside the answer. The trigger map is concrete, not magical:
atrial fibrillation / afib → CHA₂DS₂-VASc and HAS-BLED
traumatic brain injury / altered mental status → GCS
cirrhosis / decompensated cirrhosis → MELD-Na
CKD / AKI → eGFR - CKD-EPI 2021 -
strep pharyngitis → Centor / McIsaac
When the calculator runs from a chat turn rather than from a form the clinician filled in, the inputs it had to assume are flagged on the card - for example, "Age assumed 60; edit to recompute" -, and a one click handoff opens the same calculator's form with the inputs already populated, so the clinician can change anything that does not match the patient and watch the score recompute. The trigger is the diagnosis, the computation is the same deterministic code that runs from the gallery, and the result threads back into the answer with its components intact.
The point
The flashy part of clinical AI is the language. The trustworthy part is knowing exactly where the language stops and arithmetic begins. Cynaps draws that line clearly: thirty eight validated scores across eighteen specialties, computed by code that matches the published formula to the decimal, transparent down to each contributing input, cited to the paper, and ready the moment a diagnosis calls for them.
