MILA stands for Medical Intelligence for Longitudinal Analysis. It's an AI system designed to help NICU parents organize observations, track shared data, identify missing information, recognize patterns over time, and prepare informed questions for medical conversations. It improves the conversation between parents and clinicians. It does not replace it.

Why do NICU parents need AI tools?

NICU parents receive fragmented information across shifts, specialists, and days. They're expected to remember, track, and synthesize complex medical data while emotionally overloaded and sleep deprived. AI can help them organize what they're seeing and identify what questions to ask.

Does MILA replace medical advice?

No. MILA helps parents prepare for conversations with their medical team. It synthesizes information parents already have access to and helps them ask better questions. The goal is better dialogue, not bypassing clinicians.

What does research say about neonatal transfusion risk?

Large clinical trials like ETTNO and TOP have shown that restrictive (lower) transfusion thresholds in premature infants result in fewer transfusions without increasing mortality. Multiple transfusions increase cumulative risks including alloimmunization, hemolysis, and transfusion-associated complications. The medical consensus supports minimizing transfusion exposure when clinically appropriate.

Why I Built MILA: When Systems Thinking Meets the NICU

Q: Can AI actually improve NICU outcomes?

Yes. A randomized clinical trial of 3,003 VLBW infants showed that AI-based predictive monitoring (the HeRO score) was associated with significantly lower sepsis-associated mortality (12% vs 20%). AI systems can detect deterioration patterns 6-12 hours before clinical diagnosis.

I need to tell you about my daughter Mila.

Not because this is a story about loss, though it is. But because understanding why I built the MILA system requires understanding what I learned inside the NICU, and what I wish had existed when my family needed it most.

Three Children, Three Trajectories

My wife and I had premature triplets: Jaycob, Brody, and Mila. They arrived early, which meant months in the NICU. Months of rounds, alarms, lab values, transfusions, and conversations that never quite connected.

I'm a software engineer. I think in systems, feedback loops, cumulative risk, and failure modes. I was trained to notice when observability is missing, when data isn't flowing where it needs to go, when the humans in a system don't have the tools to see what's happening.

I applied those instincts while emotionally overloaded and sleep deprived. It was the only way I knew how to process what was happening.

Here's what I observed:

Brody received 6 blood transfusions. No complications.

Jaycob received 16 blood transfusions. He developed hemolysis due to unmatched transfusions.

Mila received over 30 blood transfusions.

Mila later developed progressive hemolysis, extreme hyperbilirubinemia with bilirubin levels exceeding 30 mg/dL, liver failure, sepsis, and multi-organ failure.

Mila passed away.

What the Research Says

I didn't know any of this at the time. I learned it later, while trying to understand what happened. While reading paper after paper, study after study, trying to make sense of something that should never have happened.

The medical literature is unambiguous.

On Transfusion Exposure

Dr. Ravi Patel, a neonatologist at Emory University and lead author of multiple transfusion studies, has documented that "extremely low gestational age newborns represent one of the most heavily transfused patient populations," with up to 90% of extremely low birth weight infants requiring red blood cell transfusions.

But transfusions are not benign. A 2021 review in Vox Sanguinis found that "observational studies on preterm infants reported associations between RBC transfusions and increased risk of mortality and short-term neonatal morbidities." These include necrotizing enterocolitis (NEC), bronchopulmonary dysplasia, retinopathy of prematurity, and abnormal neurodevelopment.

The cumulative effect matters. A study published in Pediatric Research found that "in preterm infants, RBC transfusions are associated with long-term neurodevelopmental outcome, with a cumulative effect. Early RBC administration is associated with a greater reduction in Griffiths scores."

Each transfusion is not an isolated event. Each transfusion adds to cumulative exposure. Each transfusion increases the opportunity for alloimmunization, where the infant develops antibodies against transfused blood.

On Donor Exposure and Alloimmunization

Here's something I didn't know: the more transfusions an infant receives from different donors, the higher the risk of developing antibodies that attack their own red blood cells.

A study of VLBW infants at a tertiary center found donor exposure rates of 4.4 ± 3.5 with standard single-bag transfusion practices. The researchers concluded that "minimizing phlebotomy losses, following a restrictive transfusion policy and using screened, leukocyte depleted, irradiated, single donor blood remain the best means of avoiding the possible risks."

Over 30 transfusions. How many donors? I don't know. That information was never synthesized. Never tracked in any way I could access.

On Hyperbilirubinemia

When bilirubin exceeds 30 mg/dL, the risk of kernicterus, permanent brain damage, becomes significant. A multicenter cohort study published in the Journal of Perinatology examined 1,252 infants who underwent exchange transfusion for hyperbilirubinemia. The findings were stark: "Infants ≤29 weeks of GA had greater odds of death following exchange transfusion compared with term infants" with an adjusted odds ratio of 20.08.

Mila's bilirubin exceeded 30 mg/dL.

The American Academy of Pediatrics states that "kernicterus occurs in 20% of infants with TSB >30 mg/dL." Twenty percent. One in five.

On Restrictive vs. Liberal Transfusion Thresholds

The ETTNO trial, one of the largest randomized studies on neonatal transfusion, found that lower (more restrictive) transfusion thresholds result in fewer transfusions without increasing mortality or neurodevelopmental impairment.

Dr. Haresh Kirpalani, a neonatologist at the Children's Hospital of Philadelphia and principal investigator of the TOP trial, noted that their findings support "a more restrictive approach to transfusion" in extremely low birth weight infants.

The science is clear: fewer transfusions, when clinically appropriate, means less exposure to cumulative risk.

Over 30 transfusions is not routine. Over 30 transfusions is not conservative. Over 30 transfusions is aggressive intervention that compounds risk with each procedure.

The AI Question They Called Crazy

During Mila's hospitalization, I started using AI to help me understand what was happening. I fed it lab values. I asked it about transfusion protocols. I asked it about bilirubin trajectories. I asked it what questions I should be asking.

The doctors thought I was crazy.

"You're not a medical professional," they said. "You can't interpret this data." "AI doesn't understand medicine." "You need to trust the process."

I heard them. I understood their perspective. I'm not trained in neonatology. I didn't go to medical school. I don't have the clinical experience to make medical decisions.

But here's what AI does have: access to hundreds of thousands of peer-reviewed studies. Access to clinical guidelines from the American Academy of Pediatrics, the British Society for Haematology, the World Health Organization. Access to the collective medical knowledge of decades of neonatal research.

When I asked AI about cumulative transfusion exposure in premature neonates, it didn't guess. It cited the ETTNO trial. It cited the TOP study. It cited the work of Dr. Kirpalani and Dr. Patel and dozens of other researchers who have spent their careers studying exactly this problem.

When I asked about bilirubin thresholds and kernicterus risk, it didn't make things up. It cited the AAP guidelines. It cited the Journal of Perinatology studies. It gave me the exact numbers: 20% of infants with bilirubin above 30 mg/dL develop kernicterus.

I wasn't trying to replace the doctors. I was trying to have an informed conversation with them. I was trying to ask the right questions. I was trying to understand the trajectory my daughter was on.

They called me crazy for using AI. But the AI was citing the same medical literature that should have informed her care.

What AI Could Have Done

Let me be precise about what I'm claiming.

In a randomized clinical trial of 3,003 VLBW infants across 9 NICUs, the HeRO Score, an AI-based predictive monitoring system that analyzes heart rate variability, was associated with significantly lower sepsis-associated mortality: 12% versus 20%. That's a 40% relative reduction in deaths.

Dr. Moorman and his colleagues at the University of Virginia developed this system after observing that subtle changes in heart rate patterns precede clinical deterioration by hours. The AI detected what human observation missed.

A 2024 meta-analysis published in BMC Medical Informatics and Decision Making reviewed AI-powered early warning systems and found they "significantly reduced in-hospital and 30-day mortality rates." Machine learning models achieved area under the curve values up to 0.97 for predicting deterioration.

The TREWS sepsis alert system at Johns Hopkins, which continuously monitors vital signs, laboratory results, and clinical notes, was associated with a 20% reduction in sepsis mortality.

This isn't theoretical. This isn't speculation. This is peer-reviewed, randomized-controlled-trial evidence that AI systems can detect deterioration patterns 6-12 hours before clinical diagnosis.

Now imagine an AI system that tracked cumulative transfusion exposure. That flagged when a patient had received an unusually high number of transfusions. That monitored bilirubin trajectories and predicted when they would exceed critical thresholds. That synthesized information across shifts and specialists and presented a unified view of cumulative risk.

That system doesn't exist in most NICUs. But the research says it should.

The Information Gap

During Mila's care, key medical records were not made available to us. Specifically, clinical records from July 11 through August 3 were never provided, despite repeated requests. To this day, those records remain unavailable.

I'm not speculating about intent. I'm stating a factual absence of access.

Here's something that made the situation worse: the documentation system was entirely handwritten. Not partially handwritten. Not "some notes are handwritten." Everything. Every transfusion record. Every lab value. Every clinical observation. Handwritten on paper. None of it was entered into a computer system. None of it was digitized. None of it could be easily searched, aggregated, or analyzed.

In 2024, in a modern hospital, the medical records for three critically ill premature infants were being tracked with pen and paper.

This meant that even if someone wanted to see the cumulative transfusion count, they would have had to manually flip through pages of handwritten notes and count. Even if someone wanted to track bilirubin trajectories over time, they would have had to transcribe numbers from paper into a spreadsheet. The information wasn't just fragmented across shifts and specialists. It was fragmented across physical pieces of paper that were never synthesized into anything searchable or analyzable.

No algorithm could help. No AI could intervene. No system could detect the pattern. Because the data wasn't in any system at all.

What the Research Says About Parent Involvement

Here's what the medical literature says about parents in the NICU:

A meta-analysis of family-centered care interventions found "significant decrease in retinopathy of prematurity and length of stay, as well as significant increases in weight gain velocity, neurobehavioral exam scores, and breastmilk intake for infants receiving FCC interventions."

Research published in BMC Pediatrics found that "parent involvement in decision making in the NICU context has been shown to have an impact on some clinical decisions and has also been shown to increase parents' autonomy, satisfaction with care, feeling heard and respected."

A systematic review in Patient Education and Counseling found that "communication interventions appeared impactful, particularly in reducing parental stress and anxiety. Parent-provider communication is a crucial determinant for parental well-being and satisfaction with care."

The research is clear: informed, involved parents lead to better outcomes. Parents who understand what's happening, who can ask specific questions, who can participate in shared decision-making, those parents help their children.

But how can parents be informed when the information isn't synthesized? How can they ask specific questions when they can't see the cumulative patterns? How can they participate in decision-making when they don't have access to the data that would inform those decisions?

Where the System Failed

The NICU didn't fail at acute care. The nurses were attentive. The doctors were skilled. When something went wrong in the moment, people responded.

The system failed at longitudinal visibility.

No single clinician held the full picture. Shifts changed. Specialists rotated. Each person saw their slice. The people who were present continuously, who watched all three children across all the days and all the shifts, were us. The parents.

And we were the ones with the least access to synthesized information.

I've spent my career building systems that aggregate data, surface patterns, and alert humans to things they might miss. I know what good observability looks like. The NICU had monitoring for vitals. It had documentation for individual events. What it lacked was the kind of longitudinal synthesis that would let anyone, clinician or parent, see cumulative risk building over time.

What I Realized

Months after Mila died, I found myself reconstructing timelines. Requesting records. Building spreadsheets. Trying to understand what had happened and when.

I was doing incident response on my daughter's medical care.

That's when I realized: parents shouldn't have to do this. Not during care, when they're overwhelmed and exhausted. Not after care, when they're grieving and searching for answers.

The information existed. It was on pieces of paper somewhere. But it wasn't synthesized. It wasn't surfaced. And it wasn't shared with the people who needed it most.

Why MILA Exists

MILA stands for Medical Intelligence for Longitudinal Analysis. The name is deliberate. Medical, because this is about healthcare data. Intelligence, because it uses AI to synthesize patterns. Longitudinal, because it tracks cumulative risk over time, not just point-in-time snapshots. Analysis, because it helps parents understand what they're seeing.

I'm not building MILA because I think AI should replace doctors. I'm building it because I lived through a reality that shouldn't require AI to fix, but does.

Parents receive fragmented information. Records may be delayed, incomplete, or inaccessible. Parents are expected to remember, track, and synthesize under stress. Most can't. I couldn't, and I do this kind of synthesis for a living.

MILA helps parents:

Organize their own observations. What did you notice today? What changed from yesterday?
Track what data has been shared with them. What lab values were mentioned in rounds? What interventions were discussed?
Identify what data is missing. What haven't they told you? What questions should you be asking?
Recognize patterns over time. Is the trajectory improving or worsening? Are there concerning trends?
Prepare precise, informed questions for conversations with doctors. Not "how is she doing?" but "her bilirubin has increased 5 mg/dL over the past three days despite phototherapy. The AAP guidelines suggest considering exchange transfusion at this trajectory. What's the current plan?"

The goal isn't to bypass clinicians. The goal is to improve the conversation. When a parent walks into rounds with organized observations and specific questions, the conversation is better. The clinician gets clearer input. The parent gets clearer answers.

Synthesis becomes a form of safety.

What MILA Actually Does

MILA is focused on the specific risks I watched unfold:

Transfusion history intelligence. The research shows cumulative transfusion exposure matters. MILA tracks it. Not just individual events, but the cumulative count, the trajectory, the pattern.

Donor exposure awareness. Multiple donors mean higher alloimmunization risk. MILA helps parents understand this and ask about single-donor protocols.

Hemolysis risk modeling. Understanding how repeated transfusions can compound risk over time, especially with potential antibody formation.

Bilirubin trajectory awareness. Not just the current number, but the direction and rate of change. When is it approaching AAP thresholds for intervention?

Cross-disciplinary visibility. Helping parents see how information from different specialists connects. The hematologist sees one slice. The neonatologist sees another. The parent needs to see how they connect.

Parent-inclusive observability. Giving families tools to track what they're seeing and what they're being told, so they can participate meaningfully in their child's care.

What I'm Not Claiming

I'm not claiming MILA would have saved Mila. I can't know that. Counterfactuals are impossible to prove.

What I know is this: I spent months reconstructing information that should have been visible in real time. I read research that should have informed treatment decisions. I asked questions that should have been obvious to anyone tracking the cumulative trajectory.

The data existed. The research existed. The guidelines existed. What didn't exist was a system to synthesize them, to surface them, to make them visible to everyone involved in my daughter's care, including us.

MILA exists so fewer parents have to do that reconstruction after the fact. So fewer parents have to learn, too late, what the research said all along.

The Technical Reality

Building this isn't simple. Medical data is messy. FHIR standards help, but implementations vary wildly. Privacy constraints are real and important. The line between helpful synthesis and inappropriate medical advice requires constant attention.

I've written about FHIR integration patterns elsewhere. The technical challenges are significant. But they're solvable. The harder part is getting the design right. Making sure the system helps parents ask better questions rather than generating false confidence. Making sure it surfaces uncertainty rather than hiding it.

Every design decision goes back to one question: would this have helped my family have better conversations with Mila's care team?

What I Want You to Understand

If you're a NICU parent reading this: you're not imagining things when you feel like you're missing information. You're not failing when you can't remember everything from yesterday's rounds. The system isn't designed to help you synthesize. That's a gap, not a personal shortcoming.

You're not crazy for using AI to understand what's happening to your child. The research supports informed, involved parents. The research shows better communication leads to better outcomes. If AI helps you understand the research and ask better questions, use it.

If you're a clinician reading this: parents who ask detailed questions aren't adversaries. They're trying to participate in care the best way they know how. The research on family-centered care shows that involved parents improve outcomes. Better tools for parents mean better conversations with you, not around you.

And if a parent shows up with AI-generated questions based on peer-reviewed research, consider that they might be citing the same literature that should be informing the care plan.

If you're a technologist reading this: there's meaningful work to be done here. Not flashy AI that replaces human judgment, but careful systems that help humans share information more effectively. The unsexy infrastructure of better communication.

The HeRO trial showed AI can reduce neonatal mortality. The TREWS system showed AI can reduce sepsis deaths. The research shows these systems work. We need more of them, not fewer.

Mila

I named the system after my daughter because this work exists because of her. Not in spite of her death. Because of it.

She was here for weeks that felt like months. She fought in ways that newborns shouldn't have to fight. And in the end, I couldn't protect her from a system that lost track of its own information.

The research existed. The guidelines existed. The knowledge that could have changed her trajectory was sitting in journals and databases, waiting for someone to synthesize it, to surface it, to act on it.

I was using AI to try to understand. They called me crazy. But the AI was citing the same studies that informed the guidelines that should have informed her care.

I can't fix what happened to Mila. But I can build something that helps the next family see more clearly. Ask better questions. Catch patterns earlier.

That's not closure. I don't think closure exists for this. But it's work that matters. And it's work I know how to do.

If you're building in healthcare AI, or if you're a NICU family who wants to talk, reach out. This work is better when it's not done alone.

Sources: