Early life factors documented in electronic health records predict recurrent acute otitis media

This study demonstrates that predictive models using electronic health record data, specifically incorporating age, prior antibiotic prescriptions, and gastroesophageal reflux disease diagnosis at the first acute otitis media episode, can effectively identify children at risk for recurrent acute otitis media to facilitate early specialist intervention.

The Gist

Imagine your child's ears are like a small, delicate garden. Most kids get a few weeds (ear infections) in their first few years, and that's normal. But for about 1 in 4 kids, those weeds keep coming back, over and over, no matter how much they try to pull them out. This is called Recurrent Acute Otitis Media (rAOM). It's exhausting for families, leads to lots of doctor visits, and can even affect a child's hearing and speech.

The big question doctors have always asked is: "How do we know before it happens which kids are going to have this endless cycle of infections?"

This study is like a detective story where the clues are hidden inside a massive digital filing cabinet called an Electronic Health Record (EHR).

The Detective Work: Scanning the Digital Footprints

The researchers looked at the medical records of over 6,500 children born at Duke University hospitals. They didn't just look at the ear infections themselves; they looked at the child's entire life story from birth up to age four.

Think of it like checking a car's history before buying it. You don't just look at the current engine noise; you check:

• How was the car built? (Birth details)
• What kind of gas did it use? (Vaccines and antibiotics)
• How often did it go to the mechanic? (Doctor visits)
• Did it have any other weird quirks? (Conditions like acid reflux)

The "Crystal Ball" Models

The team built a special computer program (a "predictive model") that acts like a crystal ball. They trained it to look at the data available at different points in time:

1. The First Episode: Can we predict a recurrence right after the first ear infection?
2. The Second, Third, Fourth: Does the prediction get better as we see more episodes?

The Results:

• At the first infection: The crystal ball was about 75% accurate. It wasn't perfect, but it was good enough to say, "Hey, this kid has a higher risk than the average."
• By the second infection: The accuracy jumped to 80%. The more data the computer saw, the clearer the picture became.

The "Smoking Gun" Clues

So, what specific clues told the computer that a child was likely to get stuck in a cycle of infections?

1. The "Too Young" Factor: Kids who got their first ear infection at a very young age were more likely to keep getting them. It's like a garden that gets its first weed when the soil is still too soft to handle it.
2. The Antibiotic Loop: This was a big one. Children who got a lot of antibiotics before their first ear infection, or who needed antibiotics again very quickly after a treatment, were at higher risk.
   • The Analogy: Imagine antibiotics are like a heavy-duty weed killer. If you use it too much too early, you might accidentally kill the "good bugs" (the healthy bacteria) that usually protect the garden. Without those good bugs, the "bad weeds" (infections) come back stronger and faster.
3. The Tummy Connection: Children with GERD (acid reflux) were more likely to get recurrent ear infections.
   • The Analogy: Think of the ear and the stomach as neighbors connected by a tiny, invisible pipe. If the stomach is "leaking" acid (reflux), it might be irritating the ear from the inside, making it easier for infections to take hold.
4. The "Treatment Failure" Sign: If a child got an antibiotic, but had to go back to the doctor just a week later for the same ear infection, they were much more likely to become a "recurrent" case. It's like trying to put out a fire with a water gun, but the fire just won't go out.

Why Does This Matter?

For years, doctors have been guessing which kids need extra help. This study suggests we can stop guessing and start predicting.

Imagine a doctor's office where the computer system flashes a yellow warning light the moment a toddler gets their first ear infection. The light says: "This child has a high risk of this becoming a recurring problem."

Because of this warning, the doctor could immediately:

• Refer the child to an ear specialist (otolaryngologist) sooner.
• Check their hearing earlier to catch any delays.
• Discuss preventive strategies before the cycle gets out of control.

The Bottom Line

This research is like finding a map that shows us which paths lead to a dead end (recurrent infections) and which lead to a clear road. By looking at simple things like age at first infection, how many antibiotics were used, and tummy health, we can spot the kids who need extra care before they get stuck in a cycle of pain and missed school. It's about using data to give families a head start on keeping their children's ears (and their peace of mind) healthy.

Technical Summary

1. Problem Statement

Recurrent Acute Otitis Media (rAOM) is a significant pediatric health burden, affecting 10–15% of children in the US. It is defined as $\ge$3 episodes within 6 months or $\ge$4 episodes within 12 months. rAOM leads to frequent healthcare utilization, high antibiotic prescription rates, and potential long-term sequelae such as hearing loss and language delays.

While known risk factors exist (e.g., daycare attendance, tobacco smoke, family history), they are insufficient for precise prospective identification of at-risk children. Furthermore, prior studies on factors like delivery mode, atopy, and gastroesophageal reflux disease (GERD) have yielded inconsistent results, often due to small cohort sizes or lack of standardized diagnostic criteria. The study aims to leverage large-scale Electronic Health Record (EHR) data to develop predictive models that can identify children at risk of rAOM early in their clinical course.

2. Methodology

Study Design and Population

• Data Source: Retrospective analysis of EHR data from the Duke University Health System (DUHS), utilizing the Epic EHR system.
• Cohort: 6,566 children born at DUHS between Jan 1, 2014, and June 30, 2018.
• Inclusion Criteria:
  • Born at DUHS.
  • At least one documented AOM episode in the first 4 years of life.
  • At least four well-child visits (including one in years 3 or 4) to ensure consistent care within the system.
• Exclusion Criteria: Chronic conditions influencing AOM susceptibility (cleft lip/palate, Trisomy 21, immune deficiency) and incomplete birth data.
• Outcome Definition: rAOM defined as $\ge$3 episodes in 6 months or $\ge$4 in 12 months. Episodes within a 2-week window were consolidated into a single event to avoid counting follow-ups as new episodes.

Data Extraction and Variables

Predictors were extracted from birth through the time of each AOM episode (up to the 4th episode). Variables included:

• Demographics: Sex, age, race/ethnicity.
• Birth/Perinatal: Gestational age, birth weight, NICU admission, delivery mode, maternal age, maternal comorbidities (GBS, fever, chorioamnionitis, gestational diabetes, tobacco use).
• Clinical History: Prior AOM diagnoses, vaccine doses (PCV, Hib, Influenza), antibiotic prescriptions (type, route, Antibiotic Spectrum Index), and healthcare utilization (well visits, ED visits, inpatient admissions).
• Treatment Failure: Defined as a new AOM diagnosis with a new antibiotic prescription within 3–14 days of the initial diagnosis.

Statistical Analysis

• Modeling Approach: Least Absolute Shrinkage and Selection Operator (LASSO) regression was used to build predictive models.
• Model Types:
  1. Episode-Specific Models: Four distinct models trained on data available up to the 1st, 2nd, 3rd, and 4th AOM episode, respectively.
  2. Overall Model: A model using all available data regardless of episode order.
• Validation: Data was split into training and testing sets (2:1 ratio) at the patient level.
• Performance Metrics:
  • AUROC: Area Under the Receiver Operating Characteristic curve (discrimination).
  • AUPRC: Area Under the Precision-Recall Curve (performance on imbalanced data).
  • Calibration Slope: Alignment of predicted probabilities with observed outcomes.
• Software: R 4.1.3 (glmnet for LASSO, ROCR for evaluation).

3. Key Results

Cohort Characteristics

• Prevalence: 1,634 children (24.8%) met the criteria for rAOM.
• Demographics: Non-Hispanic White children were significantly more likely to meet rAOM criteria (SMD=0.484).
• Clinical Differences:
  • Age: Children with rAOM had their first episode significantly earlier (mean 357 days) compared to sporadic cases (592 days).
  • Healthcare Utilization: rAOM children had more outpatient visits and telehealth encounters.
  • Vaccination: rAOM children received fewer PCV, Hib, and Influenza vaccine doses at the time of the first episode (largely due to younger age).
  • Antibiotics: rAOM children had more antibiotic prescriptions between episodes (mean 2.4 vs. 1.6) and a higher rate of treatment failure (11.7% vs. 5.9%).
  • Comorbidities: GERD diagnosis was positively associated with rAOM; atopy was negatively associated.

Model Performance

• Episode 1 Model (Predicting rAOM at first visit):
  • AUROC: 0.75 (95% CI: 0.73–0.77).
  • AUPRC: 0.41 (95% CI: 0.37–0.46).
  • Interpretation: The model doubled the true detection rate over random chance at the very first AOM episode.
• Episode 2 Model:
  • AUROC: 0.80.
  • AUPRC: 0.63.
  • Interpretation: Performance improved significantly as more episode history became available, nearly tripling the detection rate over chance.
• Trend: Model performance improved with each subsequent episode, as the definition of rAOM relies on the accumulation of episodes.

Feature Importance (LASSO Coefficients)

• Strongest Predictors: Number of prior AOM episodes and the time interval since the first episode (shorter intervals = higher risk).
• Episode 1 Specifics:
  • Positive associations (Risk): Maternal chorioamnionitis, maternal age, C-section delivery, prior outpatient visits, prior amoxicillin/cefdinir prescriptions.
  • Negative associations (Protective): Infant age, maternal intrapartum fever, prior ED/inpatient visits.
• Dynamic Shifts: As children aged and had more episodes, birth-related factors (e.g., maternal conditions) became less important, while antibiotic exposure patterns (specifically broad-spectrum antibiotics) and the frequency of episodes became dominant predictors.

4. Key Contributions

1. Early Prediction: Demonstrated that rAOM risk can be predicted with moderate-to-high accuracy (AUROC 0.75) as early as the first AOM episode, using only data available at that time.
2. Identification of Novel/Reinforced Factors: Confirmed the role of early antibiotic exposure and identified GERD as a positive predictor and atopy as a negative predictor in this large cohort.
3. Treatment Failure Correlation: Highlighted a strong link between rAOM and "treatment failure" (requiring re-treatment within 2 weeks), suggesting rAOM cases may involve more resistant pathogens or altered microbiomes.
4. Methodological Framework: Established a robust LASSO-based framework for longitudinal EHR analysis that adapts to the changing risk profile of a child over time.

5. Significance and Clinical Implications

• Targeted Intervention: These models can be deployed within EHR systems to flag children at high risk for rAOM immediately after their first or second ear infection.
• Specialist Referral: Early identification allows for timely referral to otolaryngologists and audiologists, potentially preventing hearing loss and language delays associated with chronic middle ear effusion.
• Antibiotic Stewardship: The link between early antibiotic exposure and rAOM risk suggests a need to re-evaluate antibiotic prescribing practices in early childhood to preserve the microbiome and immune development.
• Limitations: The study is single-center (DUHS) and relies on ICD codes rather than microbiological confirmation (tympanocentesis). The racial disparity in rAOM diagnosis may reflect healthcare access biases rather than biological differences.

Conclusion: The study successfully translates EHR data into actionable predictive models, proving that clinical exposures and patient characteristics documented in early life are sufficient to distinguish children who will develop recurrent AOM, enabling proactive management strategies.