Babies born too early have a brutal welcome to the world: a body that's still under construction, organs running on beta software, and - if sepsis crashes the party - an urgent need for antibiotics that were never really designed with them in mind.

Here's the problem. A 25-week preterm infant and a full-term newborn might both need gentamicin, but they're about as physiologically similar as a house cat and a lion. Their kidneys filter drugs at wildly different rates. Their body composition changes by the hour. And those aminoglycoside antibiotics? Too little means the infection wins; too much means potential kidney damage or hearing loss. Doctors have been threading this needle largely through educated guessing and frequent blood tests.

A team of researchers from Italy and Belgium just published something genuinely clever in npj Digital Medicine: a digital twin framework that simulates each individual newborn's drug metabolism, predicts how bacteria will respond and potentially evolve resistance, and then calculates the optimal antibiotic dosing schedule - all before the actual drugs are given.

What's a Digital Twin Doing in a NICU?

Digital twins started in aerospace engineering - virtual copies of jet engines that could predict when parts would fail. The medical version creates a computational clone of a patient, fed by real data, that can run "what if" scenarios faster than reality allows.

This framework combines three layers that work together like a well-choreographed medical team:

Layer 1: A physiologically-based pharmacokinetic (PBPK) model - essentially a mathematical map of how aminoglycosides travel through a neonate's body. Where does the drug go? How fast does it get there? How quickly do the kidneys clear it out?

Layer 2: An LSTM neural network - a type of AI architecture with memory, trained on real-world neonatal data - that continuously updates its estimate of each baby's glomerular filtration rate (how well the kidneys filter). This matters because kidney function changes rapidly in preterm infants and is notoriously hard to measure directly.

Layer 3: An eco-evolutionary pharmacodynamic module - and this is where it gets interesting. Instead of just modeling "drug kills bacteria," this component simulates how bacterial populations grow, die, and potentially develop resistance under antibiotic pressure. It tracks the minimum inhibitory concentration (MIC) as a moving target, not a fixed number.

Why Evolution-Aware Dosing Matters

Traditional antibiotic dosing assumes bacteria are static enemies. Hit them hard enough, they die, end of story. But bacteria evolve - sometimes even at antibiotic concentrations far below the MIC. The bugs that survive a suboptimal dose become the parents of the next generation, potentially more resistant than before.

By including an evolutionary component, the digital twin can simulate entire treatment courses and identify dosing strategies that not only kill the current infection but minimize the selection pressure that breeds resistant survivors. Think of it as playing chess against bacteria that can change the rules mid-game.

Running 1,634 Parallel Experiments

The researchers calibrated their framework using data from 1,634 actual neonates - a dataset spanning the full spectrum from extremely preterm to term newborns, including those with acute kidney injury and other complications. They then generated virtual patient cohorts to stress-test their dosing optimization under conditions you'd never ethically create in real patients.

Using model predictive control - a technique borrowed from engineering where you optimize based on predictions of near-future states - they reduced bacterial rebound during late therapy phases. The system achieved bacteriostatic drug exposure across all digital-twin neonates while keeping drug concentrations in the safe zone for most scenarios, even at elevated MIC levels.

The Catch (There's Always a Catch)

This remains a computational proof-of-concept. The framework needs prospective clinical validation - real babies, real outcomes, real comparisons to standard dosing. PBPK models in neonates are notoriously tricky because physiology changes so rapidly in the first weeks of life. Enzyme systems that don't exist at birth spin up within days; kidney function matures on timescales that make adult pharmacology seem frozen by comparison.

And there's the data question. This approach requires relatively rich physiological information to build each digital twin. In well-resourced NICUs with continuous monitoring, that's increasingly feasible. In resource-limited settings where neonatal sepsis deaths are highest, the computational infrastructure might be the last thing anyone's worrying about.

What This Points Toward

The bigger idea here extends beyond aminoglycosides. Any renally-cleared drug in any rapidly-changing patient population could potentially benefit from this kind of model-informed precision dosing. The framework architecture - PBPK plus machine learning updating plus evolutionary dynamics - represents a template that could adapt to chemotherapy in pediatric oncology, immunosuppressants in transplant patients, or anywhere else one-size-fits-all dosing collides with biological heterogeneity.

We're watching the early emergence of what a 2025 Lancet Digital Health paper calls "twin synchronization" - the continuous handshake between real patient and virtual model that keeps predictions grounded in reality as treatment unfolds.

For the tiniest, most vulnerable patients, getting drug dosing right isn't just about efficacy curves and trough levels. It's about giving them the best shot at an outcome their underdeveloped organs weren't designed to handle. Running those experiments computationally first, then applying the insights to actual care, is exactly the kind of test-before-you-fly philosophy that made digital twins valuable in aerospace. Turns out babies and jet engines have something in common after all.

References:

1. Prunella M, Romano C, Borri A, et al. Evolutionary digital twin framework for optimal aminoglycoside dosing in neonates with suspected sepsis. npj Digital Medicine. 2026. DOI: 10.1038/s41746-026-02558-w

2. Pacifici GM, Allegaert K. Developmental Pharmacokinetics of Antibiotics Used in Neonatal ICU: Focus on Preterm Infants. Biomedicines. 2023;11(3):940. PMCID: PMC10046592

3. Jahn K, et al. OptiDose: Computing the Individualized Optimal Drug Dosing Regimen Using Optimal Control. Journal of Optimization Theory and Applications. 2021. PMCID: PMC8550736

4. van der Heijden J, et al. Physiologically Based Pharmacokinetics Modeling in the Neonatal Population - Current Advances, Challenges, and Opportunities. Pharmaceutics. 2023;15(11):2579. PMCID: PMC10675397

5. Björkholm B, et al. Pervasive Selection for Clinically Relevant Resistance at Very Low Antibiotic Concentrations. PLoS Pathogens. 2023. PMID: 36627817

6. Subramanian K, et al. Model-informed precision dosing: State of the art and future perspectives. Advanced Drug Delivery Reviews. 2024. DOI: 10.1016/j.addr.2024.115436

7. Coorey G, et al. Medical digital twins: enabling precision medicine and medical artificial intelligence. The Lancet Digital Health. 2025. DOI: 10.1016/S2589-7500(25)00028-7

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.