Erasmus MC’s bet on blood-based, AI-guided transplant surveillance
A transplanted organ saves a life. Keeping it safe often means puncturing it, repeatedly. Heart recipients can expect a calendar dotted with endomyocardial biopsies in the first year alone. Each sample is a sliver of truth and a slab of inconvenience. Erasmus MC, with Deepcell’s REM-I, is attempting something that sounds modest but could be momentous: to replace much of that schedule with a simple blood draw and a great deal of computation.
The premise is disarmingly straightforward. When rejection stirs in the graft, the immune system leaves fingerprints in the blood. Deepcell’s benchtop instrument images vast numbers of circulating cells – completely label-free – and an AI model goes hunting for the faint signatures of activation and shift. It is less of a new test than a new sense: a way of seeing cell state that is continuously learning rather than forcing biology into predefined molds.
Erasmus MC’s early case study focuses on heart transplantation. The team paired routine biopsy grades (0R, 1R, 2R) with same-time peripheral blood, ran the cells through REM-I and projected millions of them into a two-dimensional map (See Figure 1). Distinct islands of morphology emerged; the biopsy grades, overlaid as densities, did not scatter randomly but formed a gradient from quiescent to inflamed with a visible “transition” belt in between. That belt is where clinical indecision typically lives. “Our early results give us hope to get a non-invasive manner of surveillance of rejection and reduce the amount of costly biopsies.” says Dr Olivier Manintveld, the cardiologist leading the effort.
The analysis did not stop at pretty plots. When the team asked which image features best separated the data-driven “morphotypes”, the answer was not crude size or texture but the model’s higher-order, learned descriptors. In other words, what matters most is what the network sees, not what the naked eye guesses. Longitudinal traces from individual patients told a similar story: some 1R episodes drifted toward the inflammatory cluster, others receded toward 0R: visual proof of a therapeutic window that biopsies, taken episodically, are apt to miss. The message was blunt about the clinical reality: “With biopsies: clinical inertia!” and, by contrast, a “window of opportunity” when a blood-based risk starts to climb.
This is not an argument for algorithmic autocracy. It is, rather, a proposal to move vigilance from the catheter lab to the clinic chair; from a rarefied, invasive snapshot to a rhythm of gentle, frequent sensing. The hoped-for protocol is plain: draw blood at follow-ups; run REM-I; combine a risk score and an explanation layer (which morphotypes are swelling; which immune classes: B cells, CD4 T cells, NK cells, monocytes are shifting); act earlier when the curve bends the wrong way. “Early results suggest that Deepcell can phenotype our immune system and be of added value both in transplantation medicine as well as in cardio-oncology… However, we have to collaborate to get further ahead.” says Dr Manintveld.
Figure 2: From blood draw to decision support – how Deepcell can help
None of this dethrones the biopsy. Pathology remains the court of record. But a court is not a traffic camera. The promise here is to turn “come back next month and we will see” into “you are drifting; let us adjust before tissue pays the price.” It is a small change in grammar, present progressive rather than past perfect, that patients might feel as fewer procedures and fewer panicked admissions.
The road ahead is properly dull: more samples, sorted cells, orthogonal assays; pragmatic thresholds by organ; explainability that clinicians can read at a glance; and the plumbing that carries results into the electronic record without fuss. Erasmus has a head start.
If this works, it will not only suit budgets and schedules. It will suit patients. Biopsies will still have their moments. But much of transplant aftercare could move from puncture to pulse-checking: quiet, routine, more humane and, crucially, earlier.
Mahyar Salek
CTO
