Clinical confidence, from a model that keeps learning

TPN2.0 was trained on a decade of real neonatal nutrition, validated in a second health system without retraining, and built to improve as practice, guidelines, and each unit's own preferences evolve.

A one-time model drifts

Most clinical models are trained once on historical data and then deployed unchanged. As prescribing patterns shift, patient populations change, and guidelines are revised, a fixed model quietly loses accuracy, and no one sees it happen.

Neonatal nutrition is an unusually moving target. Practice varies by region, there is no single universal guideline, and recommendations are revised as new evidence arrives. A dosing model that cannot move with them degrades exactly where precision matters most.

Built on a decade of real neonatal care

TPN2.0 learned from a decade of routine neonatal care, 79,790 real TPN orders across 5,913 infants, and distilled that practice into a small set of standardized, compoundable formulas. It is a physics-informed model that respects pharmacy safety limits by design, including osmolarity, maximum component concentrations, and calcium-phosphate solubility, and it keeps the physician in the loop rather than replacing them.

A decade of routine NICU data

143,000+ orders across both sites

9,330 infants across both sites

Two health systems validated without retraining

Read the peer-reviewed study in Nature Medicine →

How we know it holds up

First question, does it match how experts actually prescribe?

Across every component of the formula, TPN2.0's recommendations tracked what expert clinicians prescribe. And the same held at a second, independent hospital that had no hand in building it. We validated the model there without retraining, and the agreement stayed high.

Component-level agreement between TPN2.0 and expert clinicians at Stanford (R = 0.94) and UCSF (R = 0.91) — **Agreement with expert practice, at two hospitals.** Each dot is a nutrient. TPN2.0's distance from prescribed orders lines up with the experts' own distance, at the first hospital (R = 0.94) and at a second, independent hospital (R = 0.91, no retraining). [1]

Blinded study design and results: experts rated TPN2.0 above prescribed and random TPN — **Blinded clinicians rated TPN2.0 above best practice.** In a blinded rating of TPN2.0, the actual prescription, and a random control, TPN2.0 scored highest, 56 versus 35 for the real prescription and 20 for random. [1]

Then the harder test, would clinicians prefer it to today's best practice?

Members of the care team chart-reviewed real NICU patients and blindly rated three formulas, TPN2.0's recommendation, the actual best-practice prescription, and a random one. Across 192 comparisons, they rated TPN2.0 highest, above the prescription the patient actually received.

The model does not stand still

TPN2.0 improves along a continuous cycle. Three inputs feed model optimization, every optimized model is validated before it ships, and only then is it deployed. Then the cycle repeats.

TPN2.0

Prescriber-style adaptation

Data-partner aggregation

Guidelines updates

TPN2.0

MODEL OPTIMIZATION

Validation

Deployment

CONTINUOUS IMPROVEMENT

Prescriber-style adaptation

When a clinician adjusts a recommendation, the model can take that adjustment as the next day's starting point and move toward that prescriber's style, always inside the same safety limits. It adapts to the clinician without giving up guideline adherence.

Data-partner aggregation

De-identified datasets from research partners, like the two hospital cohorts in our study, are pooled to retrain and sharpen the model across more populations and sites. This data comes only from consented data partners, never from deployed hospitals.

Guidelines and customization

As ASPEN and institutional guidelines are revised, and as each site sets its own limits and formulary, those rules are folded in, so the model adapts to the standard of care and to the unit it serves.

Adaptation, made concrete

Here the model first suggested a higher zinc dose. A physician overrode it, and TPN2.0 carried that preference forward, holding the clinician's level for the following days rather than reverting. When the clinical picture later shifted, it moved too. The model bends toward the prescriber, within safe bounds, instead of fighting them.

After a physician lowered the zinc dose, TPN2.0 held the physician's value for the following days, then shifted with the clinical picture — **The model follows the physician's lead.** After a clinician overrode the zinc recommendation, TPN2.0 held that value for days, then adjusted as the case evolved. [1]

US Patent US20260120865A1 Systems and Methods to Assess Neonatal Health Risk and Uses Thereof View on Google Patents →

[1] Phongpreecha et al. AI-guided precision parenteral nutrition for neonatal intensive care units. Nature Medicine 31, 1882 to 1894 (2025).

409 13 St. Ste. 600 Oakland, CA 94612, USA

View the Takeoff41, Inc. Financial Conflict of Interest (FCOI) policy.