Back to blog

AI and Health

AI anomaly detection in pediatric growth charts: ML for early diagnosis

Pediatricians review the chart once a year. What if AI flagged anomalies after every measurement? Percentile drift, channel crossing, and slow growth detection via ML.

Çocuk Gelişim Scientific Board (Prof. Dr. Bülent Bayraktar)May 26, 2026 4 min read

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

ScoreRecommendation
0-20Normal, continue tracking
21-50Mild alert — re-measure in 3 months
51-80Suggest pediatrician appointment
81-100Urgent 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:

  1. Z-score + percentile computed automatically
  2. Recharts visualization — Neyzi 2008 reference bands overlay
  3. PDF report — to share with pediatricians
  4. 🔄 AI anomaly detection (planned, Q3 2026)
  5. 🔄 Personalized growth projection — 12-month forecast from historical trend
  6. 🔄 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:

  1. Sensitivity ≥ 90% — must catch 90% of real pathology
  2. Specificity ≥ 70% — must not over-alarm and create unnecessary anxiety
  3. Explainability — must justify why it alerted
  4. Clinical validation study — n=500+ prospective cohort
  5. 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.

Frequently asked questions

Who is "AI anomaly detection in pediatric growth charts: ML for early diagnosis" for?

It is written for families, coaches and clinicians who need a clear educational summary before deciding whether a pediatric evaluation is needed.

Does this article replace a pediatrician?

No. It is educational content. Diagnosis, treatment and urgent medical concerns should be handled by qualified clinicians.

What is the main takeaway?

Pediatricians review the chart once a year. What if AI flagged anomalies after every measurement? Percentile drift, channel crossing, and slow growth detection via ML.

When should families seek clinical advice?

Families should seek advice when growth velocity slows, percentiles change rapidly, puberty timing is unusual, symptoms persist, or nutrition concerns are present.

How should this content be used with calculators?

Use article context together with serial measurements and calculator warnings; do not make decisions from a single number.

#artificial-intelligence#ML#growth-chart#anomaly#prediction

⚕️ Medical disclaimer

The information in this article is for educational purposes only and does not constitute medical advice. For decisions about your child's growth, please consult a pediatrician or pediatric endocrinologist.