A child grows over years. Pediatricians see them every 6-12 months — that's few data points to detect emerging pathology. AI-based continuous monitoring between measurements can fill the gap.
Classic follow-up vs continuous AI follow-up
Classic
- Parents measure at home or at yearly check-ups in clinic
- Pediatrician reviews the chart vs. Neyzi 2008 or WHO curves: "p25, good"
- The next review is a year later — percentile may have already dropped, unnoticed
AI-augmented
- Parents log each measurement in the app
- ML model instantly assesses percentile change, growth velocity, expected trajectory
- Alert: "2-percentile-band drop in the last 6 months — pediatrician consult recommended."
What the ML pipeline does
Step 1: Normalize data
Each measurement (age, height, weight, BMI, head circumference) → Neyzi 2008 or WHO z-score. The child's position relative to same-age, same-sex peers is standardized.
Step 2: Trend analysis
- Linear regression over last N measurements → growth velocity (cm/yr)
- Percentile change over 6/12/24 months
- Channel crossing — p50 → p25 band shift
Step 3: Anomaly scoring
Statistical z-test or machine learning:
- Isolation Forest — distance from population center
- LSTM autoencoder — recognizes unexpected patterns
- Gradient Boosting (XGBoost) — pathology-risk classification
Output: 0-100 anomaly score.
Step 4: Clinical recommendation
| Score | Recommendation |
|---|---|
| 0-20 | Normal, continue tracking |
| 21-50 | Mild alert — re-measure in 3 months |
| 51-80 | Suggest pediatrician appointment |
| 81-100 | Urgent pediatric endocrinology referral |
Which anomalies can it catch?
A. Channel crossing (percentile shift)
A child tracking at p50 who drops to p15 in 12 months:
- Early sign of GH deficiency
- Hypothyroidism
- Celiac disease (especially with BMI drop)
- Psychosocial short stature
AI catches this pattern early.
B. Slowed growth velocity
In mid-childhood (4-10 yrs), <4 cm/year is pathological. Families may not notice; AI computes velocity automatically.
C. Extremely rapid growth
8+ cm/yr + early pubertal signs → precocious puberty alarm.
D. BMI explosion
Z-score rises +1 in 6 months → obesity precursor; +2 → urgent intervention.
E. Height-weight mismatch
Height p50, weight p3 → malabsorption, celiac, IBD, anorexia nervosa suspicion.
Current system architecture (Çocuk Gelişim platform)
On our platform today:
- ✅ Z-score + percentile computed automatically
- ✅ Recharts visualization — Neyzi 2008 reference bands overlay
- ✅ PDF report — to share with pediatricians
- 🔄 AI anomaly detection (planned, Q3 2026)
- 🔄 Personalized growth projection — 12-month forecast from historical trend
- 🔄 Predictive alert — automatic notification when statistically anomalous
Clinical validation — ethics and evidence
ML alarm systems are classified as Clinical Decision Support (CDSS):
- FDA Class II medical device (USA)
- CE-MDR Class IIa (Europe)
- TİTCK approval (Turkey — pediatric CDSS examples still rare)
Ethical requirements:
- Sensitivity ≥ 90% — must catch 90% of real pathology
- Specificity ≥ 70% — must not over-alarm and create unnecessary anxiety
- Explainability — must justify why it alerted
- Clinical validation study — n=500+ prospective cohort
- Legal liability — alert → suggests pediatrician; AI does not diagnose
Use cases — real world
Case 1: Celiac disease
8-year-old boy. Height p50, weight tracking at p25. AI flagged a 20% percentile drop in 6 months + low appetite logs. Referred to pediatrician — anti-tTG IgA positive. 12 months on a gluten-free diet, growth normalized.
Case 2: Constitutional delay (false positive)
12-year-old boy, stable p10. AI alerted for absent pubertal signs. Endocrinology diagnosed constitutional delay — observation sufficient. False positive, but family checkup brought reassurance.
Case 3: GH deficiency
6-year-old girl. Height dropped from p10 to p3 in 12 months. AI scored high anomaly. Endocrinology + low IGF-1 + GH stimulation testing confirmed diagnosis. Treatment → p25 within 12 months.
FAQ
What if AI sends a false alarm?
Discuss with pediatrician. AI is screening, clinicians diagnose. High sensitivity at specificity 70-80% means 1-2 of 5 alerts may be false positives — an acceptable tradeoff.
Does the algorithm use my child's data to train models?
On our platform it's opt-in. Default: no. If you choose to contribute anonymized data, you can do so in profile settings — KVKK-compliant, revocable.
Will my pediatrician accept an AI-flagged report?
Most modern pediatricians welcome more data = better decisions. Some may resist "AI alert" framing. Show them numerical trends instead: "Percentile dropped 15% in 6 months." Science-first phrasing is more palatable to all.
Is there Turkish population validation?
Not yet — pilot studies start in 2026. Our platform's target: IRB-approved n=500 cohort in Q2 2027.
Bottom line
AI-powered growth-chart anomaly detection can catch signals that classic follow-up misses. Our platform currently has rule-based anomaly detection (percentile drops, slow velocity) active. ML version is in beta for Q3 2026.
Sign up free, record measurements, and become one of the first beneficiaries of AI-based follow-up. Today you can already use charts + PDF report + Excel export.