diagcalc 3.2.4 → 5.0.0

package/lib/diagcalc-core.js

@@ -1,27 +1,40 @@
 (function (root, factory) {
   if (typeof module === "object" && module.exports) {
-    module.exports = factory();
+    module.exports = factory(require("./diagcalc-presentation"));
     return;
   }
 
-  root.DiagcalcCore = factory();
-}(typeof globalThis !== "undefined" ? globalThis : this, () => {
-  function validateInputs({ tp, fp, fn, tn, preTestProb }) {
+  root.DiagcalcCore = factory(root.DiagcalcPresentation);
+}(/** @type {any} */ (typeof globalThis !== "undefined" ? globalThis : this), (presentation) => {
+  /**
+   * @typedef {import("./diagcalc-types").DiagnosticInput} DiagnosticInput
+   * @typedef {import("./diagcalc-types").CalculationOptions} CalculationOptions
+   */
+  /** @param {unknown} input Confusion matrix counts and probability in percent. */
+  function validateInputs(input) {
+    if (!input || typeof input !== "object") {
+      return { valid: false, message: "Enter a confusion matrix and pre-test probability." };
+    }
+    const { tp, fp, fn, tn, preTestProb } = /** @type {DiagnosticInput} */ (input);
     const inputs = [tp, fp, fn, tn];
     if (inputs.some((value) => Number.isNaN(value) || value < 0)) {
       return { valid: false, message: "Enter non-negative integers for all confusion matrix cells." };
     }
 
-    if (inputs.some((value) => !Number.isInteger(value))) {
+    if (inputs.some((value) => !Number.isSafeInteger(value))) {
       return { valid: false, message: "Use whole numbers for TP, FP, FN, and TN." };
     }
 
-    if (Number.isNaN(preTestProb)) {
+    if (!Number.isSafeInteger(tp + fp + fn + tn)) {
+      return { valid: false, message: "The total must be a safe whole number (at most 9007199254740991)." };
+    }
+
+    if (!Number.isFinite(preTestProb)) {
       return { valid: false, message: "Enter a valid pre-test probability." };
     }
 
-    if (preTestProb < 0 || preTestProb >= 100) {
-      return { valid: false, message: "Pre-test probability must be at least 0% and below 100%." };
+    if (preTestProb < 0 || preTestProb > 100) {
+      return { valid: false, message: "Pre-test probability must be between 0% and 100%, inclusive." };
     }
 
     if (tp + fn === 0) {
@@ -32,18 +45,15 @@
       return { valid: false, message: "Specificity is indeterminate: add cases without disease (TN or FP)." };
     }
 
-    if (tp + fp === 0) {
-      return { valid: false, message: "PPV is indeterminate: at least one positive test result is required." };
-    }
-
-    if (tn + fn === 0) {
-      return { valid: false, message: "NPV is indeterminate: at least one negative test result is required." };
-    }
-
     return { valid: true };
   }
 
-  function calculateMetrics({ tp, fp, fn, tn, preTestProb }) {
+  /** @param {DiagnosticInput} input @param {CalculationOptions} [opts] */
+  function calculateResults(input, opts) {
+    if (!validateInputs(input).valid) return null;
+    if (opts && opts.continuityCorrection && !["auto", "always", "never"].includes(opts.continuityCorrection)) return null;
+    const { tp, fp, fn, tn, preTestProb } = input;
+    const continuityMode = normaliseContinuityMode(opts && opts.continuityCorrection);
     const totals = {
       diseased: tp + fn,
       nonDiseased: tn + fp,
@@ -56,71 +66,83 @@
     const specificity = tn / totals.nonDiseased;
     const ppv = tp / totals.positives;
     const npv = tn / totals.negatives;
-    const lrPositive = calculateRatio(sensitivity, 1 - specificity);
-    const lrNegative = calculateRatio(1 - sensitivity, specificity);
+    const lrPositive = calculateRatio(sensitivity, fp / totals.nonDiseased);
+    const lrNegative = calculateRatio(fn / totals.diseased, specificity);
     const preTestOdds = calculateOdds(preTest);
     const postTestOddsPositive = multiplyOdds(preTestOdds, lrPositive);
     const postTestOddsNegative = multiplyOdds(preTestOdds, lrNegative);
     const postTestProbPositive = probabilityFromOdds(postTestOddsPositive);
     const postTestProbNegative = probabilityFromOdds(postTestOddsNegative);
 
-    return {
+    const lrPositiveCI = Number.isNaN(lrPositive) ? null : calcLogRatioCI(tp, totals.diseased, fp, totals.nonDiseased, { continuityCorrection: continuityMode });
+    const lrNegativeCI = Number.isNaN(lrNegative) ? null : calcLogRatioCI(fn, totals.diseased, tn, totals.nonDiseased, { continuityCorrection: continuityMode });
+    const postTestPositiveCI = Number.isNaN(postTestProbPositive) ? null : calcPostTestCI(preTest, lrPositiveCI);
+    const postTestNegativeCI = Number.isNaN(postTestProbNegative) ? null : calcPostTestCI(preTest, lrNegativeCI);
+    const dor = calcDOR(tp, fp, fn, tn, { continuityCorrection: continuityMode });
+    const nns = (sensitivity > 0 && preTest > 0) ? 1 / (sensitivity * preTest) : Infinity;
+
+    const results = {
       sensitivity: {
-        label: "Sensitivity",
         value: sensitivity,
         ci: calcWilsonInterval(tp, totals.diseased),
-        note: buildSensitivityNote(sensitivity),
       },
       specificity: {
-        label: "Specificity",
         value: specificity,
         ci: calcWilsonInterval(tn, totals.nonDiseased),
-        note: buildSpecificityNote(specificity),
       },
       ppv: {
-        label: "Positive predictive value (PPV)",
         value: ppv,
         ci: calcWilsonInterval(tp, totals.positives),
-        note: "Probability of disease given a positive result.",
       },
       npv: {
-        label: "Negative predictive value (NPV)",
         value: npv,
         ci: calcWilsonInterval(tn, totals.negatives),
-        note: "Probability of no disease given a negative result.",
+      },
+      dor: {
+        value: dor.value,
+        ci: dor.ci,
+      },
+      numberNeededToScreen: {
+        value: nns,
       },
       lrPositive: {
-        label: "Positive likelihood ratio (LR+)",
         value: lrPositive,
-        formatter: formatLikelihood,
-        note: interpretLRPositive(lrPositive),
+        ci: lrPositiveCI,
       },
       lrNegative: {
-        label: "Negative likelihood ratio (LR-)",
         value: lrNegative,
-        formatter: formatLikelihood,
-        note: interpretLRNegative(lrNegative),
+        ci: lrNegativeCI,
       },
       preTestProbability: {
-        label: "Pre-test probability",
         value: preTest,
-        note: "Estimated starting point before the test result.",
       },
       postTestPositive: {
-        label: "Post-test probability (positive result)",
         value: postTestProbPositive,
-        note: "Updated probability of disease after a positive test.",
+        ci: postTestPositiveCI,
       },
       postTestNegative: {
-        label: "Post-test probability (negative result)",
         value: postTestProbNegative,
-        note: "Updated probability of disease after a negative test.",
+        ci: postTestNegativeCI,
       },
     };
+    return Object.fromEntries(Object.entries(results).map(([key, metric]) => [key, {
+      ...metric,
+      unit: ["dor", "lrPositive", "lrNegative"].includes(key) ? "ratio" : key === "numberNeededToScreen" ? "count" : "probability",
+      interpretation: { id: key, params: { value: metric.value } },
+    }]));
+  }
+
+  // Compatibility adapter: new integrations may consume calculateResults()
+  // directly; all prose and formatter selection live in presentation.
+  /** @param {DiagnosticInput} input @param {CalculationOptions} [opts] */
+  function calculateMetrics(input, opts = {}) {
+    const results = calculateResults(input, opts);
+    return results && presentation.presentResults(results, { formatLikelihood, formatNNS }, opts && opts.locale);
   }
 
+  /** @param {number} successes @param {number} total */
   function calcWilsonInterval(successes, total) {
-    if (total === 0) {
+    if (!Number.isSafeInteger(total) || !Number.isSafeInteger(successes) || successes < 0 || successes > total || total <= 0) {
       return null;
     }
 
@@ -134,7 +156,436 @@
     return { lower, upper };
   }
 
+  // 95% CI for a ratio of two binomial proportions on the log scale.
+  // Used for LR+: (x1, n1) = (TP, diseased), (x2, n2) = (FP, non-diseased).
+  // Used for LR-: (x1, n1) = (FN, diseased), (x2, n2) = (TN, non-diseased).
+  // Simel DL, Samsa GP, Matchar DB. J Clin Epidemiol 1991;44(8):763–770.
+  // Continuity correction (+0.5 on each cell) applied when any cell is 0.
+  function calcLogRatioCI(x1, n1, x2, n2, opts) {
+    if (![x1, n1, x2, n2].every(Number.isSafeInteger) || x1 < 0 || x2 < 0 || n1 <= 0 || n2 <= 0 || x1 > n1 || x2 > n2) return null;
+    const mode = normaliseContinuityMode(opts && opts.continuityCorrection);
+    let a = x1;
+    let b = n1 - x1;
+    let c = x2;
+    let d = n2 - x2;
+    const hasZero = a === 0 || b === 0 || c === 0 || d === 0;
+
+    if (mode === "always" || (mode === "auto" && hasZero)) {
+      a += 0.5;
+      b += 0.5;
+      c += 0.5;
+      d += 0.5;
+    }
+    // "never" mode: leave cells as-is — the existing checks below catch the
+    // cases where the maths actually fails (p1 or p2 = 0).
+
+    const tn1 = a + b;
+    const tn2 = c + d;
+    const p1 = a / tn1;
+    const p2 = c / tn2;
+
+    if (p1 === 0 || p2 === 0) {
+      return null;
+    }
+
+    const logRatio = Math.log(p1 / p2);
+    const seLog = Math.sqrt((1 - p1) / a + (1 - p2) / c);
+
+    if (!Number.isFinite(seLog) || !Number.isFinite(logRatio)) {
+      return null;
+    }
+
+    const z = 1.96;
+    return {
+      lower: Math.exp(logRatio - z * seLog),
+      upper: Math.exp(logRatio + z * seLog),
+    };
+  }
+
+  function normaliseContinuityMode(value) {
+    if (value === "always" || value === "never") return value;
+    return "auto";
+  }
+
+  // Cohen's kappa for inter-rater agreement on a binary outcome.
+  // Inputs: the four cells of the rater1×rater2 2×2 table.
+  // Returns { value, ci, observed, expected, n, interpretation } or null on
+  // empty input. Marginal-adjusted asymptotic variance for the CI; Landis & Koch (1977) bins
+  // for the qualitative interpretation.
+  function calcCohenKappa(input) {
+    if (!input || typeof input !== "object") return null;
+    const { bothPos, only1Pos, only2Pos, bothNeg } = input;
+    const cells = [bothPos, only1Pos, only2Pos, bothNeg];
+    if (cells.some((v) => !Number.isSafeInteger(v) || v < 0)) return null;
+    const N = bothPos + only1Pos + only2Pos + bothNeg;
+    if (N === 0 || !Number.isSafeInteger(N)) return null;
+
+    const po = (bothPos + bothNeg) / N;
+    const r1Pos = (bothPos + only1Pos) / N;
+    const r2Pos = (bothPos + only2Pos) / N;
+    const r1Neg = (only2Pos + bothNeg) / N;
+    const r2Neg = (only1Pos + bothNeg) / N;
+    const pe = r1Pos * r2Pos + r1Neg * r2Neg;
+
+    if (pe >= 1 - 1e-12) {
+      return {
+        value: NaN,
+        ci: null,
+        observed: po,
+        expected: pe,
+        n: N,
+        interpretation: "Expected agreement is 1; kappa is undefined.",
+      };
+    }
+
+    const kappa = (po - pe) / (1 - pe);
+    // Unweighted asymptotic variance including estimated marginal frequencies.
+    // Reference implementation: statsmodels.stats.inter_rater.cohens_kappa.
+    const rows = [r1Pos, r1Neg];
+    const cols = [r2Pos, r2Neg];
+    const diagonal = [bothPos / N, bothNeg / N];
+    const termA = diagonal.reduce((sum, p, i) =>
+      sum + p * (1 - (rows[i] + cols[i]) * (1 - kappa)) ** 2, 0);
+    const termB = (1 - kappa) ** 2 * (
+      (only1Pos / N) * (cols[0] + rows[1]) ** 2 +
+      (only2Pos / N) * (cols[1] + rows[0]) ** 2
+    );
+    const termC = (kappa - pe * (1 - kappa)) ** 2;
+    const variance = Math.max(0, (termA + termB - termC) / ((1 - pe) ** 2 * N));
+    const seKappa = Math.sqrt(variance);
+    const z = 1.96;
+    return {
+      value: kappa,
+      ci: { lower: kappa - z * seKappa, upper: kappa + z * seKappa },
+      observed: po,
+      expected: pe,
+      n: N,
+      interpretation: interpretKappa(kappa),
+    };
+  }
+
+  function interpretKappa(value) {
+    return presentation.noteFor("interpretKappa", value);
+  }
+
+  // ROC reconstruction from a list of (cutoff, TP, FP, FN, TN) rows, one per
+  // threshold of a continuous test. Returns the sorted points on the (FPR, TPR)
+  // plane, trapezoidal AUC including the (0,0) and (1,1) anchors, and the index
+  // of the row with the highest Youden's J = sens + spec − 1.
+  /** @param {unknown} rows @param {{direction?: string}} [opts] */
+  function validateRocRows(rows, opts = {}) {
+    const errors = [];
+    if (!Array.isArray(rows) || rows.length === 0 || rows.length > 500) {
+      return { valid: false, errors: [{ row: null, message: "Use between one and 500 complete cutoff rows." }] };
+    }
+    let totals = null;
+    rows.forEach((row, index) => {
+      const validation = validateInputs(row && { ...row, preTestProb: 50 });
+      if (!validation.valid) {
+        errors.push({ row: index + 1, message: validation.message });
+        return;
+      }
+      const current = [row.tp + row.fn, row.fp + row.tn];
+      if (totals && (totals[0] !== current[0] || totals[1] !== current[1])) {
+        errors.push({ row: index + 1, message: "All cutoffs must use the same diseased and non-diseased cohorts." });
+      }
+      totals = totals || current;
+      if (row.cutoff !== null && row.cutoff !== undefined && !Number.isFinite(row.cutoff)) {
+        errors.push({ row: index + 1, message: "Cutoff must be a finite number or left blank." });
+      }
+    });
+    if (errors.length) return { valid: false, errors };
+    const sorted = rows.map((row, index) => ({ ...row, index }))
+      .sort((a, b) => a.fp - b.fp || a.tp - b.tp);
+    for (let i = 1; i < sorted.length; i += 1) {
+      if (sorted[i].tp < sorted[i - 1].tp) {
+        errors.push({ row: sorted[i].index + 1, message: "Sensitivity must not decrease as the false-positive rate increases." });
+      }
+    }
+    const direction = opts.direction === "lower" ? "lower" : "higher";
+    const cutoffs = sorted.filter((row) => Number.isFinite(row.cutoff));
+    for (let i = 1; i < cutoffs.length; i += 1) {
+      const previous = cutoffs[i - 1];
+      const current = cutoffs[i];
+      const changed = current.tp !== previous.tp || current.fp !== previous.fp;
+      const ordered = direction === "higher" ? current.cutoff < previous.cutoff : current.cutoff > previous.cutoff;
+      if (changed && !ordered) {
+        errors.push({ row: current.index + 1, message: `Cutoffs must follow the selected ${direction}-score-positive direction.` });
+      }
+    }
+    return { valid: errors.length === 0, errors };
+  }
+
+  function calculateROC(rows, opts) {
+    if (!validateRocRows(rows, opts).valid) return null;
+    const points = rows.map((row) => {
+      const sens = row.tp / (row.tp + row.fn);
+      const spec = row.tn / (row.tn + row.fp);
+      return {
+        ...row,
+        cutoff: row.cutoff ?? null,
+        provenance: row.provenance || (row.synthetic ? "simulated" : "observed"),
+        sens,
+        spec,
+        fpr: row.fp / (row.tn + row.fp),
+        youden: sens + spec - 1,
+      };
+    }).sort((a, b) => a.fpr - b.fpr || a.sens - b.sens);
+
+    const augmented = [
+      { fpr: 0, sens: 0 },
+      ...points,
+      { fpr: 1, sens: 1 },
+    ];
+    let auc = 0;
+    for (let i = 1; i < augmented.length; i += 1) {
+      const dx = augmented[i].fpr - augmented[i - 1].fpr;
+      if (dx <= 0) continue;
+      auc += dx * (augmented[i].sens + augmented[i - 1].sens) / 2;
+    }
+
+    let optimalIndex = -1;
+    let maxYouden = -Infinity;
+    for (let i = 0; i < points.length; i += 1) {
+      if (points[i].youden > maxYouden) {
+        maxYouden = points[i].youden;
+        optimalIndex = i;
+      }
+    }
+
+    return {
+      points,
+      simulated: points.some((point) => point.provenance !== "observed"),
+      auc,
+      optimalIndex,
+      optimalPoint: optimalIndex >= 0 ? points[optimalIndex] : null,
+    };
+  }
+
+  // Standard normal CDF Φ(x). Abramowitz & Stegun 7.1.26 polynomial approximation
+  // of erf; max error ~1.5e-7 — fine for ROC scaffolding.
+  /** @param {number} x */
+  function stdNormalCdf(x) {
+    if (Number.isNaN(x)) return NaN;
+    if (!Number.isFinite(x)) return x > 0 ? 1 : 0;
+    const sign = x < 0 ? -1 : 1;
+    const ax = Math.abs(x) / Math.SQRT2;
+    const t = 1 / (1 + 0.3275911 * ax);
+    const y = 1 - (((((1.061405429 * t - 1.453152027) * t) + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax);
+    return 0.5 * (1 + sign * y);
+  }
+
+  // Inverse standard normal Φ⁻¹(p) via Acklam's rational approximation.
+  // Max relative error ~1.15e-9 in (0, 1).
+  /** @param {number} p */
+  function invStdNormal(p) {
+    if (!(p > 0 && p < 1)) {
+      if (p === 0) return -Infinity;
+      if (p === 1) return Infinity;
+      return NaN;
+    }
+    const a = [-3.969683028665376e+1, 2.209460984245205e+2, -2.759285104469687e+2, 1.383577518672690e+2, -3.066479806614716e+1, 2.506628277459239e+0];
+    const b = [-5.447609879822406e+1, 1.615858368580409e+2, -1.556989798598866e+2, 6.680131188771972e+1, -1.328068155288572e+1];
+    const c = [-7.784894002430293e-3, -3.223964580411365e-1, -2.400758277161838e+0, -2.549732539343734e+0, 4.374664141464968e+0, 2.938163982698783e+0];
+    const d = [7.784695709041462e-3, 3.224671290700398e-1, 2.445134137142996e+0, 3.754408661907416e+0];
+    const pLow = 0.02425;
+    const pHigh = 1 - pLow;
+    let q, r;
+    if (p < pLow) {
+      q = Math.sqrt(-2 * Math.log(p));
+      return (((((c[0] * q + c[1]) * q + c[2]) * q + c[3]) * q + c[4]) * q + c[5]) /
+        ((((d[0] * q + d[1]) * q + d[2]) * q + d[3]) * q + 1);
+    }
+    if (p <= pHigh) {
+      q = p - 0.5;
+      r = q * q;
+      return (((((a[0] * r + a[1]) * r + a[2]) * r + a[3]) * r + a[4]) * r + a[5]) * q /
+        (((((b[0] * r + b[1]) * r + b[2]) * r + b[3]) * r + b[4]) * r + 1);
+    }
+    q = Math.sqrt(-2 * Math.log(1 - p));
+    return -(((((c[0] * q + c[1]) * q + c[2]) * q + c[3]) * q + c[4]) * q + c[5]) /
+      ((((d[0] * q + d[1]) * q + d[2]) * q + d[3]) * q + 1);
+  }
+
+  // Synthesize a scaffold of ROC cutoffs from a single observed 2×2.
+  // Assumes a binormal (equal-variance) score model: scores in non-diseased
+  // ~ N(0,1), in diseased ~ N(d,1). The observed sens/spec pin down d via
+  // d = Φ⁻¹(sens) + Φ⁻¹(spec). Then we sweep `count` cutoffs at evenly spaced
+  // FPRs in (0, 1), round the implied counts to the observed P/N totals.
+  // Returns an array of { cutoff, tp, fp, fn, tn, synthetic: true } — the
+  // cutoff is the implied normalized score (Φ⁻¹(1-FPR)) rounded to 2 dp.
+  function generateSyntheticRocPoints(observed, count) {
+    if (!validateInputs(observed && { ...observed, preTestProb: 50 }).valid) return [];
+    const { tp, fp, fn, tn } = observed;
+    if (!Number.isInteger(tp) || !Number.isInteger(fp) || !Number.isInteger(fn) || !Number.isInteger(tn)) return [];
+    const P = tp + fn;
+    const N = tn + fp;
+    if (P <= 0 || N <= 0) return [];
+    const sens = tp / P;
+    const spec = tn / N;
+    // Edge: a perfect (or pathological) observed point has Φ⁻¹(0) = -∞.
+    // Nudge sens/spec away from 0 and 1 so d stays finite.
+    const eps = 1 / (2 * Math.max(P, N));
+    const sClamp = Math.min(1 - eps, Math.max(eps, sens));
+    const pClamp = Math.min(1 - eps, Math.max(eps, spec));
+    const d = invStdNormal(sClamp) + invStdNormal(pClamp);
+    const n = Math.max(2, Math.min(20, Number.isInteger(count) ? count : 7));
+    // Span (edge, 1 − edge) inclusively so the scaffolded points reach the
+    // bottom-left and top-right of the ROC plot. Previous fence-post sampling
+    // (i / (n + 1)) topped out around FPR 0.83 and bottomed out around 0.17,
+    // leaving the corners empty.
+    const edge = 0.025;
+    const rows = [];
+    for (let i = 0; i < n; i += 1) {
+      const fpr = n === 1 ? 0.5 : edge + (1 - 2 * edge) * i / (n - 1);
+      const c = invStdNormal(1 - fpr);
+      const tprImplied = 1 - stdNormalCdf(c - d);
+      const tpI = Math.max(0, Math.min(P, Math.round(tprImplied * P)));
+      const fpI = Math.max(0, Math.min(N, Math.round(fpr * N)));
+      rows.push({
+        cutoff: Number(c.toFixed(2)),
+        tp: tpI,
+        fp: fpI,
+        fn: P - tpI,
+        tn: N - fpI,
+        synthetic: true,
+        provenance: "simulated",
+      });
+    }
+    return rows;
+  }
+
+  // Diagnostic odds ratio with log-normal CI.
+  // DOR = (TP·TN) / (FP·FN); SE(log DOR) = sqrt(1/TP + 1/FP + 1/FN + 1/TN).
+  // Continuity correction (+0.5 to every cell) if any cell is 0.
+  function calcDOR(tp, fp, fn, tn, opts) {
+    if (![tp, fp, fn, tn].every((value) => Number.isSafeInteger(value) && value >= 0) || !Number.isSafeInteger(tp + fp + fn + tn) || tp + fp + fn + tn === 0) return { value: NaN, ci: null };
+    const mode = normaliseContinuityMode(opts && opts.continuityCorrection);
+    let a = tp;
+    let b = fp;
+    let c = fn;
+    let d = tn;
+    const hasZero = a === 0 || b === 0 || c === 0 || d === 0;
+    if (mode === "always" || (mode === "auto" && hasZero)) {
+      a += 0.5;
+      b += 0.5;
+      c += 0.5;
+      d += 0.5;
+    }
+    // never + hasZero → DOR is 0 or ∞ depending on which cell is zero; the CI is undefined.
+    const rawValue = mode === "never" ? (tp * tn) / (fp * fn) : (a * d) / (b * c);
+    if (!Number.isFinite(rawValue) || rawValue <= 0) {
+      return { value: rawValue, ci: null };
+    }
+    const seLog = Math.sqrt(1 / a + 1 / b + 1 / c + 1 / d);
+    const logDor = Math.log(rawValue);
+    const z = 1.96;
+    return {
+      value: rawValue,
+      ci: {
+        lower: Math.exp(logDor - z * seLog),
+        upper: Math.exp(logDor + z * seLog),
+      },
+    };
+  }
+
+  function interpretDOR(value) {
+    return presentation.noteFor("interpretDOR", value);
+  }
+
+  // Surface common biases that the 2x2 alone can signal.
+  // Returns an array of plain-text warnings; empty array if nothing flagged.
+  /** @param {DiagnosticInput} input */
+  function buildBiasWarnings({ tp, fp, fn, tn, preTestProb }) {
+    const warnings = [];
+    const nDiseased = tp + fn;
+    const nNonDiseased = tn + fp;
+    const total = nDiseased + nNonDiseased;
+
+    if (nDiseased > 0 && nDiseased < 30) {
+      warnings.push(`Small diseased group (n=${nDiseased}). Sensitivity and likelihood-ratio estimates are imprecise — interpret 95% CIs carefully.`);
+    }
+    if (nNonDiseased > 0 && nNonDiseased < 30) {
+      warnings.push(`Small non-diseased group (n=${nNonDiseased}). Specificity and likelihood-ratio estimates are imprecise — interpret 95% CIs carefully.`);
+    }
+
+    if (total > 0 && Number.isFinite(preTestProb)) {
+      const studyPrev = nDiseased / total;
+      const userPrev = preTestProb / 100;
+      if (Math.abs(userPrev - studyPrev) > 0.20) {
+        warnings.push(
+          `Study prevalence (${(studyPrev * 100).toFixed(1)}%) differs substantially from your pre-test probability (${preTestProb.toFixed(1)}%). The PPV and NPV cards reflect the 2x2's prevalence; use the post-test probabilities for your patient.`
+        );
+      }
+    }
+
+    return warnings;
+  }
+
+  // Pauker DG, Kassirer JP. The threshold approach to clinical decision making.
+  // N Engl J Med 1980;302(20):1109–1117.
+  //
+  // Given a clinician-stated treatment threshold Pt (the post-test probability
+  // at which the expected utility of treating equals that of not treating) and
+  // the diagnostic test's LR+ and LR−, return the two pre-test thresholds that
+  // bracket the "testing is useful" zone:
+  //
+  //   testingThreshold       (P_low):  below this pre-test, even a positive
+  //                                    result keeps post-test < Pt → don't test
+  //   testTreatmentThreshold (P_high): above this pre-test, even a negative
+  //                                    result keeps post-test ≥ Pt → just treat
+  //
+  // Derivation: solving post-odds = pre-odds × LR for the pre-test probability
+  // that maps to Pt on the post-test scale gives
+  //     P = Pt / (Pt + (1 − Pt) × LR)
+  /** @param {{treatmentThreshold: number, lrPositive: number, lrNegative: number}} opts */
+  function calculateThresholds(opts) {
+    if (!opts) return null;
+    const Pt = opts.treatmentThreshold;
+    if (!Number.isFinite(Pt) || Pt <= 0 || Pt >= 1) {
+      return null;
+    }
+
+    if (!(opts.lrPositive > 1) || !(opts.lrNegative >= 0 && opts.lrNegative < 1)) return null;
+
+    function thresholdFromLR(lr) {
+      if (Number.isNaN(lr)) return null;
+      if (lr === Infinity) return 0;
+      if (lr === 0) return 1;
+      if (!Number.isFinite(lr) || lr < 0) return null;
+      return Pt / (Pt + (1 - Pt) * lr);
+    }
+
+    return {
+      treatmentThreshold: Pt,
+      testingThreshold: thresholdFromLR(opts.lrPositive),
+      testTreatmentThreshold: thresholdFromLR(opts.lrNegative),
+    };
+  }
+
+  // Propagate the LR CI to the post-test probability scale via the Bayes' update
+  // Monotonic Bayes transformation preserves endpoints; the prior is fixed.
+  function calcPostTestCI(preTestProbability, lrCi) {
+    if (!lrCi) {
+      return null;
+    }
+    if (!Number.isFinite(preTestProbability) || preTestProbability < 0 || preTestProbability > 1) {
+      return null;
+    }
+
+    const preOdds = calculateOdds(preTestProbability);
+    const lower = probabilityFromOdds(multiplyOdds(preOdds, lrCi.lower));
+    const upper = probabilityFromOdds(multiplyOdds(preOdds, lrCi.upper));
+
+    if (Number.isNaN(lower) || Number.isNaN(upper)) {
+      return null;
+    }
+    return { lower, upper };
+  }
+
+  /** @param {number} numerator @param {number} denominator */
   function calculateRatio(numerator, denominator) {
+    if (!Number.isFinite(numerator) || !Number.isFinite(denominator) || numerator < 0 || denominator < 0) return NaN;
     if (numerator === 0 && denominator === 0) {
       return NaN;
     }
@@ -147,7 +598,9 @@
     return numerator / denominator;
   }
 
+  /** @param {number} probability */
   function calculateOdds(probability) {
+    if (!Number.isFinite(probability) || probability < 0 || probability > 1) return NaN;
     if (probability === 1) {
       return Infinity;
     }
@@ -157,17 +610,21 @@
     return probability / (1 - probability);
   }
 
+  /** @param {number} odds @param {number} ratio */
   function multiplyOdds(odds, ratio) {
-    if (odds === 0 || ratio === 0) {
-      return 0;
-    }
-    if (!Number.isFinite(odds) || !Number.isFinite(ratio)) {
-      return Infinity;
-    }
+    if (typeof odds !== "number" || typeof ratio !== "number" || Number.isNaN(odds) || Number.isNaN(ratio) || odds < 0 || ratio < 0) return NaN;
+    if ((odds === 0 && ratio === Infinity) || (ratio === 0 && odds === Infinity)) return NaN;
     return odds * ratio;
   }
 
+  /** @param {number} probability */
+  function chainedPreTestProbability(probability) {
+    return Number.isFinite(probability) && probability >= 0 && probability <= 1 ? probability * 100 : NaN;
+  }
+
+  /** @param {number} odds */
   function probabilityFromOdds(odds) {
+    if (typeof odds !== "number" || Number.isNaN(odds) || odds < 0) return NaN;
     if (odds === Infinity) {
       return 1;
     }
@@ -184,6 +641,7 @@
     return formatPercentage(value);
   }
 
+  /** @param {number} value */
   function formatPercentage(value) {
     if (!Number.isFinite(value)) {
       return "—";
@@ -191,6 +649,7 @@
     return `${(value * 100).toFixed(1)}%`;
   }
 
+  /** @param {unknown} value */
   function normaliseDecimal(value) {
     if (typeof value !== "string") {
       return "";
@@ -198,6 +657,7 @@
     return value.replace(",", ".").trim();
   }
 
+  /** @param {unknown} value */
   function safeParseInt(value) {
     if (typeof value !== "string") {
       return NaN;
@@ -209,73 +669,60 @@
     if (!/^\d+$/.test(trimmed)) {
       return NaN;
     }
-    return parseInt(trimmed, 10);
+    const parsed = Number(trimmed);
+    return Number.isSafeInteger(parsed) ? parsed : NaN;
+  }
+
+  /** @param {unknown} value */
+  function parseProbability(value) {
+    const normalised = normaliseDecimal(value);
+    if (!/^(?:\d+(?:\.\d*)?|\.\d+)$/.test(normalised)) return NaN;
+    const parsed = Number(normalised);
+    return Number.isFinite(parsed) && parsed >= 0 && parsed <= 100 ? parsed : NaN;
   }
 
+  /** @param {number} value */
   function formatLikelihood(value) {
-    if (!Number.isFinite(value)) {
+    if (value === Infinity) {
       return "∞";
     }
+    if (!Number.isFinite(value)) return "—";
     if (value === 0) {
       return "0";
     }
     return value >= 10 ? value.toFixed(1) : value.toFixed(2);
   }
 
+  /** @param {number} value */
+  function formatNNS(value) {
+    if (!Number.isFinite(value)) return "—";
+    if (value <= 0) return "—";
+    return String(Math.ceil(value));
+  }
+
+  function interpretNNS(value) {
+    return presentation.noteFor("interpretNNS", value);
+  }
+
+  /** @param {number} value @param {number} min @param {number} max */
   function clamp(value, min, max) {
     return Math.min(Math.max(value, min), max);
   }
 
   function buildSensitivityNote(value) {
-    if (value >= 0.9) {
-      return "Captures most cases with disease. Useful for screening.";
-    }
-    if (value >= 0.7) {
-      return "Moderate sensitivity: consider alongside clinical context.";
-    }
-    return "Low sensitivity: consider additional tests to reduce false negatives.";
+    return presentation.noteFor("buildSensitivityNote", value);
   }
 
   function buildSpecificityNote(value) {
-    if (value >= 0.9) {
-      return "Few false positives. Suitable for confirming diagnoses.";
-    }
-    if (value >= 0.7) {
-      return "Moderate specificity: confirm with other laboratory or clinical data.";
-    }
-    return "Low specificity: beware of false positives and their impact on treatment decisions.";
+    return presentation.noteFor("buildSpecificityNote", value);
   }
 
   function interpretLRPositive(value) {
-    if (!Number.isFinite(value)) {
-      return "LR+ is very high: a positive result virtually confirms the disease.";
-    }
-    if (value >= 10) {
-      return "LR+ >= 10 indicates strong evidence in favour of disease.";
-    }
-    if (value >= 5) {
-      return "Moderate LR+: substantially increases the probability of disease.";
-    }
-    if (value >= 2) {
-      return "Low LR+: limited gain; combine with other data.";
-    }
-    return "LR+ close to 1: a positive result barely changes the probability of disease.";
+    return presentation.noteFor("interpretLRPositive", value);
   }
 
   function interpretLRNegative(value) {
-    if (!Number.isFinite(value)) {
-      return "LR- is infinite: a negative result does not reduce the probability of disease.";
-    }
-    if (value <= 0.1) {
-      return "LR- <= 0.1 indicates strong evidence against disease.";
-    }
-    if (value <= 0.2) {
-      return "Moderate LR-: reduces the probability of disease.";
-    }
-    if (value <= 0.5) {
-      return "Low LR-: limited impact; consider further evaluation.";
-    }
-    return "LR- close to 1: a negative result does not rule out disease.";
+    return presentation.noteFor("interpretLRNegative", value);
   }
 
   function buildProbabilityBar(value, width) {
@@ -286,23 +733,41 @@
   }
 
   return {
+    buildBiasWarnings,
     buildProbabilityBar,
     buildSensitivityNote,
     buildSpecificityNote,
+    calcCohenKappa,
+    calcDOR,
+    calcLogRatioCI,
+    calcPostTestCI,
     calcWilsonInterval,
     calculateMetrics,
+    calculateResults,
+    chainedPreTestProbability,
     calculateOdds,
+    calculateROC,
     calculateRatio,
+    calculateThresholds,
     clamp,
     formatLikelihood,
+    formatNNS,
     formatPercentage,
     formatValue,
+    generateSyntheticRocPoints,
+    interpretDOR,
+    interpretKappa,
     interpretLRNegative,
     interpretLRPositive,
+    interpretNNS,
+    invStdNormal,
     multiplyOdds,
     normaliseDecimal,
     probabilityFromOdds,
     safeParseInt,
+    stdNormalCdf,
     validateInputs,
+    validateRocRows,
+    parseProbability,
   };
 }));