Does it work?

Evaluation

The synthetic cohort generator hides prescribing cascades inside 600 patients by sampling real CAUSES and TREATED_BY edges, and mixes in the confounders a real record has: drugs prescribed for a genuine diagnosis, deliberate pairs started together, and drugs started in the wrong order. The ground truth never enters the graph. Each row adds one graph signal.

0.75

precision

0.96

recall

0.84

of 227 deliberate pairs wrongly flagged

Counterfactual replay

If Domino had been at the prescriber’s desk

Every one of the cohort’s 2,443 prescriptions is re-checked as the record stood the day it was written (only earlier prescriptions and diagnoses count). When the check stops a true cascade, that drug and everything that followed from it are removed from the rest of the replay.

75%

of 534 cascades stopped at the first downstream prescription (401)

510

prescriptions that might never have been written

0.95

of cascade stops were true cascades (22 were not)

1/210

deliberate companions (a PPI with an NSAID) stopped as a cascade; 7 stopped for completing a dangerous combination

522 of 534 cascades got at least a caution · 20.9 stops per 100 prescriptions in a cohort seeded with cascades · 794 cautions · 19.4 s

Ablation: what each graph signal adds

632 hidden cascades · 2154 candidate links found by the graph · threshold 0.25 · 22.5 s on FalkorDB

Graph links only

Any CAUSES → TREATED_BY link between two of the patient’s drugs.

Precision

0.30

Recall

1.00

0.46

tp 632 · fp 1463 · fn 0 · deliberate pairs flagged 227

+ time order

The second drug must start after the first.

Precision

0.44

Recall

1.00

0.61

tp 630 · fp 810 · fn 2 · deliberate pairs flagged 200

+ onset window

Score by how common the effect is and whether the gap fits its usual onset.

Precision

0.59

Recall

1.00

0.74

tp 630 · fp 434 · fn 2 · deliberate pairs flagged 200

+ other reasons

Down-weight when the second drug has its own diagnosis, or the effect was already diagnosed.

Precision

0.58

Recall

0.96

0.72

tp 606 · fp 446 · fn 26 · deliberate pairs flagged 226

Domino (full)

Adds documented patterns and routes deliberate pairs (gastroprotection, potassium) aside.

Precision

0.75

Recall

0.96

0.84

tp 606 · fp 204 · fn 26 · deliberate pairs flagged 0

Recall by pattern

Graph-inferred (no documented pattern)416/424
NSAID → high blood pressure → antihypertensive17/29
ACE inhibitor → dry cough → cough suppressant23/23
Diuretic → urinary incontinence → bladder medication21/22
Calcium channel blocker → ankle swelling → diuretic21/21
Alpha-1 blocker → dizziness → vestibular sedative19/19
Metformin → diarrhoea → loperamide18/18
Thiazide → gout → gout treatment12/13
SSRI or SNRI → insomnia → sleep medication11/12
Bisphosphonate → indigestion → acid suppression10/10
Anticholinergic → urinary retention → alpha blocker7/9
Bladder antimuscarinic → memory problems → dementia medication9/9
Gabapentinoid → ankle swelling → loop diuretic6/6
Benzodiazepine → memory problems → dementia medication3/3
Verapamil → constipation → laxative3/3
Benzodiazepine → agitation or delirium → antipsychotic3/3
Colchicine → diarrhoea → loperamide3/3
Cholinesterase inhibitor → urinary incontinence → bladder antimuscarinic2/2
Corticosteroid → high blood sugar → diabetes medication2/2
Antipsychotic → parkinsonism → antiparkinsonian0/1

Read this honestly

This tests the reasoning, not clinical truth. The generator and the engine share one knowledge graph, so graph-level recall is high by construction. The point is the jump in precision as each signal is added.
The misses are real ones: cascades where the second drug also had a genuine diagnosis on record, which Domino correctly treats as less certain.
Clinical validation would need real prescribing data, for example a prescription sequence symmetry analysis like Doherty et al. (2025), and pharmacist review of flagged records.
The knowledge graph is hand-curated for a hackathon: frequency bands and onset windows are approximate.
The prevention replay counts a cascade as caught only on a hard stop at the first downstream prescription; cautions are counted separately. “Avoided” means the replay removes that drug and every drug the ground truth says followed from it.

Evaluation

0.75

precision

0.96

recall

0.84

of 227 deliberate pairs wrongly flagged

Counterfactual replay

If Domino had been at the prescriber’s desk

75%

of 534 cascades stopped at the first downstream prescription (401)

510

prescriptions that might never have been written

0.95

of cascade stops were true cascades (22 were not)

1/210

deliberate companions (a PPI with an NSAID) stopped as a cascade; 7 stopped for completing a dangerous combination

522 of 534 cascades got at least a caution · 20.9 stops per 100 prescriptions in a cohort seeded with cascades · 794 cautions · 19.4 s

Ablation: what each graph signal adds

632 hidden cascades · 2154 candidate links found by the graph · threshold 0.25 · 22.5 s on FalkorDB

Graph links only

Any CAUSES → TREATED_BY link between two of the patient’s drugs.

Precision

0.30

Recall

1.00

0.46

tp 632 · fp 1463 · fn 0 · deliberate pairs flagged 227

+ time order

The second drug must start after the first.

Precision

0.44

Recall

1.00

0.61

tp 630 · fp 810 · fn 2 · deliberate pairs flagged 200

+ onset window

Score by how common the effect is and whether the gap fits its usual onset.

Precision

0.59

Recall

1.00

0.74

tp 630 · fp 434 · fn 2 · deliberate pairs flagged 200

+ other reasons

Down-weight when the second drug has its own diagnosis, or the effect was already diagnosed.

Precision

0.58

Recall

0.96

0.72

tp 606 · fp 446 · fn 26 · deliberate pairs flagged 226

Domino (full)

Adds documented patterns and routes deliberate pairs (gastroprotection, potassium) aside.

Precision

0.75

Recall

0.96

0.84

tp 606 · fp 204 · fn 26 · deliberate pairs flagged 0

Recall by pattern

Graph-inferred (no documented pattern)416/424
NSAID → high blood pressure → antihypertensive17/29
ACE inhibitor → dry cough → cough suppressant23/23
Diuretic → urinary incontinence → bladder medication21/22
Calcium channel blocker → ankle swelling → diuretic21/21
Alpha-1 blocker → dizziness → vestibular sedative19/19
Metformin → diarrhoea → loperamide18/18
Thiazide → gout → gout treatment12/13
SSRI or SNRI → insomnia → sleep medication11/12
Bisphosphonate → indigestion → acid suppression10/10
Anticholinergic → urinary retention → alpha blocker7/9
Bladder antimuscarinic → memory problems → dementia medication9/9
Gabapentinoid → ankle swelling → loop diuretic6/6
Benzodiazepine → memory problems → dementia medication3/3
Verapamil → constipation → laxative3/3
Benzodiazepine → agitation or delirium → antipsychotic3/3
Colchicine → diarrhoea → loperamide3/3
Cholinesterase inhibitor → urinary incontinence → bladder antimuscarinic2/2
Corticosteroid → high blood sugar → diabetes medication2/2
Antipsychotic → parkinsonism → antiparkinsonian0/1

Read this honestly

This tests the reasoning, not clinical truth. The generator and the engine share one knowledge graph, so graph-level recall is high by construction. The point is the jump in precision as each signal is added.
The misses are real ones: cascades where the second drug also had a genuine diagnosis on record, which Domino correctly treats as less certain.
Clinical validation would need real prescribing data, for example a prescription sequence symmetry analysis like Doherty et al. (2025), and pharmacist review of flagged records.
The knowledge graph is hand-curated for a hackathon: frequency bands and onset windows are approximate.
The prevention replay counts a cascade as caught only on a hard stop at the first downstream prescription; cautions are counted separately. “Avoided” means the replay removes that drug and every drug the ground truth says followed from it.