translational analytics

Forecasting adverse events from multiomic longitudinal trajectories

An AI model reads each marker's time series (not single draws) across dose cohorts, produces a per-patient early-warning score at every visit, and names the markers × dose levels driving each flag. Synthetic cohort — no real patient data.

signal markersdecoy / noise markersAE onsetearly-warning flag
Signal strength0.8×
Noise level1.0×
Cohort seed
seed 42
60 patients · 4 dose cohorts · 11 visits (Day 0–63) · trajectory features → dose-aware logistic model
Held-out ROC-AUC
0.97
20 test patients, patient-level split
Median lead time
12d
flag → recorded AE onset
Flagged ≥7d early
89%
of 9 held-out AE patients
Any early flag
100%
flagged before onset
Index patient

One patient, tracked visit by visit

Marker trajectories · log-fold vs Day 0
-0.40.51.42.33.2ramp window0371421283542495663AE D45study daylog-fold elevation
IL-6IFN-γTNF-αCRPActivated CD8+IL-10(decoy)
Early-warning score · updated each visit
0.00.30.50.81.0lead time0371421283542495663threshold 0.32AE D45flag D35study dayrisk score
Flagged 10 days early. Risk crossed threshold on Day 35; AE recorded Day 45.
Driver attribution · log-odds contribution at flag (Day 35)
Activated CD8+
+1.71
IL-6
+1.34
CRP
+0.98
TNF-α
+0.84
NK activation
+0.80
IFN-γ
+0.68
Agent summary · plain-language
P20 (Very High-dose cohort) is flagged on Day 35, 10 days before the adverse event recorded on Day 45. The risk score is driven mainly by Activated CD8+ (up sharply over the last ~2 weeks), CRP (up sharply over the last ~2 weeks), TNF-α (a sudden step up). These are the same cytokine and immune-activation markers that, in this Very High-dose cohort, began climbing during the 22-day window ahead of onset — so the model is reacting to genuine longitudinal movement, not a single abnormal draw. Recommendation: increase monitoring cadence and review for early intervention.
Population

Which markers, and how the signal scales with dose

Top biomarker drivers · model importance (signal vs decoy)
IL-6
1.90
CRP
0.96
Activated CD8+
0.92
Albumin
0.79
TNF-α
0.77
NK activation
0.72
IL-10
0.64
Ferritin
0.64
Red = markers with true injected signal. The strongest ones surface at the top of the ranking — the model was never told which markers carry signal, and the decoys settle below them.
Lead-time distribution · held-out AE patients
1
0
1
5
4
10
2
15
20
1
25
lead time before AE onset (days)
Dose-dependence of the signal
Adverse-event rate
13%
Low
47%
Medium
60%
High
80%
Very High
Signal ramp lead (days before onset)
13d
Low
15d
Medium
17d
High
20d
Very High
Peak signal elevation (log-fold)
1.07
Low
1.09
Medium
2.18
High
2.67
Very High
Higher dose cohorts have higher AE rates, steeper marker elevation, and signals that begin rising earlier before onset — the dose-dependent structure the model exploits to warn sooner in the cohorts that need it most.