Human Lung Digital Twins Provide Paired Controls for Alteplase Evaluation

Each treated lung is compared with a forecast

A drug experiment using donated human lungs faces a difficult control problem: untreated organs can differ substantially from the treated organs. Researchers at University Health Network and collaborating academic institutions developed digital twins of lungs on ex vivo lung perfusion (EVLP), a system that maintains and measures an isolated lung outside the body. Their model forecasts a lung's untreated trajectory from its own baseline data, providing a paired computational reference for a physical drug experiment.

The study reviewed 1,000 consecutive Toronto EVLP cases from 2008–2024. After excluding 49 with incomplete documentation, 951 cases contributed to modeling and analyses, with different sample counts for individual modalities. Sixteen alteplase-treated cases were left out of model training. An additional 45 clinical EVLP cases from the same center in 2024–2025 provided an independent test dataset.

A multimodal model forecasts short-term organ function

The framework combines established lung-mechanics calculations with machine learning. High-resolution ventilation measurements inform gated recurrent unit models, a recurrent neural-network architecture for time-series data. XGBoost, a gradient-boosted decision-tree method, forecasts hourly measurements. Other components incorporate biochemical, radiographic and molecular features.

Static forecasts use baseline information, while dynamic forecasts incorporate newly available measurements. For the drug comparison, first-hour baseline data are used to predict what would have happened without treatment during the next 2 hours. This keeps post-treatment measurements out of the untreated forecast used as the comparator.

In the development evaluation, the authors reported mean absolute percentage errors of 4–13% for hourly physiological forecasts and edema predictions within approximately 30 mL of observed values. They compared forecast errors with variability in historical control cohorts. The separate 45-lung test set showed errors consistent with development cross-validation for physiological, biochemical and image-derived features. The study also generated 50 simulated profiles from neighboring historical cases as an illustrative cohort; those profiles are distinct from the independent clinical test lungs.

Alteplase supplies an experimental use case

During EVLP, lungs with suspected pulmonary embolism received 20 mg alteplase after their first-hour assessment. Of the 16 treated cases, 2 were excluded because perfusion ended before sufficient follow-up. The analyzed group comprised 6 lungs subsequently judged suitable for transplantation and 8 judged unsuitable.

Pulmonary arterial pressure was the main efficacy endpoint, reflecting the intended reduction in vascular obstruction. Edema was assessed as a short-term safety-related endpoint. Each measured treated trajectory was compared with its corresponding untreated digital forecast. A conventional analysis also used a randomly selected historical untreated cohort of six lungs.

For the six suitable lungs, the conventional cohort comparison did not find a significant pressure difference at either 1 or 2 hours after treatment. The paired digital-twin comparison found lower pulmonary arterial pressure at 2 hours, with P = 0.031. Edema was not significantly different at 2 hours; the reported P = 0.063 does not establish an edema reduction.

The eight unsuitable lungs did not show a pressure-lowering response relative to their digital twins. This negative subgroup result limits the drug interpretation: the observed benefit was confined to the suitable-lung group. It also shows why a single pooled efficacy statement would lose information about the condition of the tested organs.

Forecast accuracy and context remain consequential

The paired approach reduces between-organ variation, but its untreated reference is still predicted. An apparent treatment difference therefore depends on calibration of that forecast and the relevance of the historical data. The 45-lung validation cohort supplies a separate test within the originating center, while broader external validation remains necessary.

The model focuses on the first 3 hours of EVLP, and some molecular modalities include fewer than 100 cases. Much of the dataset reflects transplantation assessment, including its specific imaging and measurement practices. The model also represents an isolated lung rather than whole-body drug distribution or off-target effects.

The authors recommend retaining a small number of physical control lungs as studies begin to use digital controls. They also envisage extending the forecast to longer perfusion intervals and incorporating treatment-specific data for other interventions. Those extensions would address uses beyond this short-term alteplase comparison.