inou/lib/loinc.go

package lib

import (
	"encoding/json"
	"fmt"
	"log"
	"strings"
)

// LoincInfo holds official LOINC data from loinc_lab.
type LoincInfo struct {
	Code      string `db:"loinc_num"`
	LongName  string `db:"long_name"`
	ShortName string `db:"short_name"`
	Component string `db:"component"`
	System    string `db:"system"`
	Property  string `db:"property"`
}

// LoincGet returns official LOINC info from loinc_lab.
func LoincGet(code string) *LoincInfo {
	var results []LoincInfo
	RefQuery("SELECT loinc_num, long_name, short_name, component, system, property FROM loinc_lab WHERE loinc_num = ?", []any{code}, &results)
	if len(results) > 0 {
		return &results[0]
	}
	return nil
}

// LoincAbbr derives a short abbreviation from loinc_lab.short_name.
// Examples: "Hgb Bld-mCnc" → "Hgb", "Neutrophils/leuk NFr Bld" → "Neut"
func LoincAbbr(info *LoincInfo) string {
	if info == nil {
		return ""
	}
	s := info.ShortName
	// Take first token (before space), strip trailing /... for differentials
	if i := strings.IndexByte(s, ' '); i > 0 {
		s = s[:i]
	}
	if i := strings.IndexByte(s, '/'); i > 0 {
		s = s[:i]
	}
	// Common abbreviation overrides
	overrides := map[string]string{
		"Neutrophils": "Neut", "Lymphocytes": "Lymph", "Monocytes": "Mono",
		"Eosinophils": "Eos", "Basophils": "Baso", "Platelets": "PLT",
		"Leukocytes": "WBC", "Erythrocytes": "RBC", "Hemoglobin": "Hgb",
		"Hematocrit": "Hct", "Glucose": "Glu", "Creatinine": "Cr",
		"Sodium": "Na", "Potassium": "K", "Chloride": "Cl",
		"Calcium": "Ca", "Albumin": "Alb", "Phosphate": "Phos",
		"Magnesium": "Mg",
	}
	// Match on component first word for overrides
	comp := info.Component
	if i := strings.IndexAny(comp, "/."); i > 0 {
		comp = comp[:i]
	}
	if abbr, ok := overrides[comp]; ok {
		return abbr
	}
	// Truncate long abbreviations
	if len(s) > 8 {
		s = s[:8]
	}
	return s
}

// LoincLookup resolves a hospital test name to a LOINC code.
// Checks cache first; on miss, uses Gemini expand → search → pick → cache.
// Returns empty string if lookup fails (no candidates, LLM error, etc).
func LoincLookup(name, specimen, unit string) string {
	// 1. Check cache
	cacheKey := strings.ToLower(name + "|" + specimen + "|" + unit)
	var cached []struct {
		LoincCode string `db:"loinc_code"`
	}
	RefQuery("SELECT loinc_code FROM loinc_cache WHERE cache_key = ?", []any{cacheKey}, &cached)
	if len(cached) > 0 {
		return cached[0].LoincCode
	}

	// 2. No Gemini key = can't do LLM lookup
	if GeminiKey == "" {
		return ""
	}

	// 3. Expand + search + pick
	lookupUnit := unit
	if lookupUnit == "%" {
		lookupUnit = "percentage"
	}

	tokens := loincTokenize(name + " " + specimen + " " + lookupUnit)
	if expanded, err := loincExpand(name, specimen, lookupUnit); err == nil {
		tokens = expanded
	}

	candidates := loincSearch(tokens)

	// Filter: if unit is %, drop count codes (NCnc)
	if unit == "%" {
		var filtered []LoincInfo
		for _, c := range candidates {
			if c.Property != "NCnc" {
				filtered = append(filtered, c)
			}
		}
		if len(filtered) > 0 {
			candidates = filtered
		}
	}

	if len(candidates) == 0 {
		log.Printf("loinc-lookup: no candidates for %q", name)
		return ""
	}

	// 4. LLM pick
	code, lname, err := loincPick(name, specimen, lookupUnit, candidates)
	if err != nil {
		log.Printf("loinc-lookup: pick failed for %q: %v", name, err)
		return ""
	}

	// 5. Cache
	RefExec(`INSERT OR REPLACE INTO loinc_cache (cache_key, input_name, input_specimen, input_unit, loinc_code, loinc_name, confidence)
		VALUES (?, ?, ?, ?, ?, ?, 'llm')`, cacheKey, name, specimen, unit, code, lname)

	return code
}

// --- internal helpers ---

func loincTokenize(s string) []string {
	s = strings.ToLower(s)
	for _, c := range []string{",", ";", "(", ")", "[", "]", "/", "-", ".", ":"} {
		s = strings.ReplaceAll(s, c, " ")
	}
	var tokens []string
	seen := map[string]bool{}
	for _, t := range strings.Fields(s) {
		if len(t) < 2 || seen[t] {
			continue
		}
		tokens = append(tokens, t)
		seen[t] = true
	}
	return tokens
}

func loincSearch(tokens []string) []LoincInfo {
	if len(tokens) == 0 {
		return nil
	}
	type entry struct {
		c    LoincInfo
		hits int
	}
	entries := map[string]*entry{}

	for _, t := range tokens {
		pattern := "%" + t + "%"
		query := "SELECT loinc_num, long_name, short_name, system, component, property FROM loinc_lab WHERE " +
			"LOWER(long_name) LIKE ? OR LOWER(short_name) LIKE ? OR LOWER(component) LIKE ?"
		var results []LoincInfo
		RefQuery(query, []any{pattern, pattern, pattern}, &results)
		for _, c := range results {
			if e, ok := entries[c.Code]; ok {
				e.hits++
			} else {
				entries[c.Code] = &entry{c: c, hits: 1}
			}
		}
	}

	minHits := 2
	if len(tokens) <= 1 {
		minHits = 1
	}

	type scored struct {
		c     LoincInfo
		score int
	}
	var scoredResults []scored
	for _, e := range entries {
		if e.hits < minHits {
			continue
		}
		s := e.hits * 100
		compLen := len(e.c.Component)
		if compLen > 0 && compLen < 50 {
			s += 50 - compLen
		}
		if !strings.Contains(e.c.Component, "/") {
			s += 20
		}
		scoredResults = append(scoredResults, scored{e.c, s})
	}

	for i := range scoredResults {
		for j := i + 1; j < len(scoredResults); j++ {
			if scoredResults[j].score > scoredResults[i].score {
				scoredResults[i], scoredResults[j] = scoredResults[j], scoredResults[i]
			}
		}
	}

	var top []LoincInfo
	for i, s := range scoredResults {
		if i >= 30 {
			break
		}
		top = append(top, s.c)
	}
	return top
}

func loincExpand(name, specimen, unit string) ([]string, error) {
	prompt := fmt.Sprintf(`Given a lab test, return search terms to find it in the LOINC database.
LOINC uses formal medical terminology (e.g. "Leukocytes" not "White Blood Cells", "Erythrocytes" not "Red Blood Cells", "Oxygen" not "O2" or "pO2").

Lab test:
  Name: %s
  Specimen: %s
  Unit: %s

Return a JSON object: {"terms": ["term1", "term2", ...]}
Include: the LOINC component name, specimen system code (e.g. Bld, BldA, BldC, BldV, Ser/Plas, Urine), and any synonyms that might appear in LOINC long names.
Keep it to 3-6 terms. JSON only.`, name, specimen, unit)

	resp, err := CallGemini(prompt)
	if err != nil {
		return nil, err
	}

	var result struct {
		Terms []string `json:"terms"`
	}
	if err := json.Unmarshal([]byte(resp), &result); err != nil {
		return nil, fmt.Errorf("parse expand response: %w", err)
	}

	var terms []string
	seen := map[string]bool{}
	for _, t := range result.Terms {
		t = strings.ToLower(strings.TrimSpace(t))
		if t != "" && !seen[t] {
			terms = append(terms, t)
			seen[t] = true
		}
	}
	for _, t := range loincTokenize(name + " " + specimen) {
		if !seen[t] {
			terms = append(terms, t)
			seen[t] = true
		}
	}
	return terms, nil
}

func loincPick(name, specimen, unit string, candidates []LoincInfo) (string, string, error) {
	var lines []string
	for i, c := range candidates {
		display := c.LongName
		display = strings.ReplaceAll(display, "/100 ", "percentage of ")
		display = strings.ReplaceAll(display, "fraction", "percentage")
		lines = append(lines, fmt.Sprintf("%d. %s — %s [System: %s]", i+1, c.Code, display, c.System))
	}

	prompt := fmt.Sprintf(`You are a clinical laboratory informatics system. Given a lab test, pick the BEST matching LOINC code from the candidate list.

Lab test:
  Name: %s
  Specimen: %s
  Unit: %s

Candidates:
%s

Return ONLY a JSON object: {"pick": <number>, "loinc": "<code>", "name": "<long name>"}
Pick the candidate that best matches the test name, specimen type, and unit. If none match well, pick the closest.
JSON only, no explanation.`, name, specimen, unit, strings.Join(lines, "\n"))

	resp, err := CallGemini(prompt)
	if err != nil {
		return "", "", fmt.Errorf("LLM call failed: %w", err)
	}

	var result struct {
		Pick  int    `json:"pick"`
		Loinc string `json:"loinc"`
		Name  string `json:"name"`
	}
	if err := json.Unmarshal([]byte(resp), &result); err != nil {
		return "", "", fmt.Errorf("parse LLM response: %w", err)
	}

	if result.Loinc == "" && result.Pick > 0 && result.Pick <= len(candidates) {
		result.Loinc = candidates[result.Pick-1].Code
		result.Name = candidates[result.Pick-1].LongName
	}
	for _, c := range candidates {
		if c.Code == result.Loinc {
			return result.Loinc, c.LongName, nil
		}
	}
	if result.Pick > 0 && result.Pick <= len(candidates) {
		c := candidates[result.Pick-1]
		return c.Code, c.LongName, nil
	}
	return "", "", fmt.Errorf("LLM returned %q (pick %d) — not in %d candidates", result.Loinc, result.Pick, len(candidates))
}