Some harm only exists in the order of events.
A drug’s side effect gets misread as a new disease and treated with a second drug. Each prescription looks correct on its own, and they are often written months apart by different doctors. Senetika reads the sequence.
One cascade, as it appears in a record
Illustrative composite — not a real patientNo alert fires anywhere along this line. Amlodipine and furosemide do not interact, so an interaction checker stays silent. Both drugs pass an appropriateness screen on their own. The only evidence that something went wrong is the spacing between the dates.
The problem
Medication safety software looks at a snapshot. Cascades happen over time.
Why the tools miss it
Interaction checkers compare the drugs a patient holds concurrently and fire when two co-present agents conflict. A prescribing cascade is not a conflict — the two drugs have no pharmacological quarrel at all. Appropriateness screens such as Beers and STOPP judge one drug at a time against patient characteristics, and in a cascade each drug can pass independently.
The signal is not in the drug list. It is in the order and spacing of events, and no deployed system reasons over that.
Why the clinicians miss it
The two prescriptions are frequently written by two different prescribers, weeks or months apart. Neither sees the whole arc. The second prescriber inherits a symptom, not a history, and treats the symptom in front of them.
Nobody makes a mistake they could reasonably have noticed. That is precisely the argument for giving the job to software: this is a memory and bookkeeping problem across time and across clinicians, which is the kind of problem computers are for.
How often
Until 2025 nobody had agreed which cascades are the wrong ones.
An international expert panel published the first consensus list of 65 potentially inappropriate prescribing cascades in 2025. In 2026 the same group measured every one of them across 2,297,942 older adults. Those two papers turned a clinical intuition into a defined, quantified target — and that is what makes a detector possible.
| Finding | Rate | Population |
|---|---|---|
| Cascades occurring in 1% or more of at-risk patients per year | 49 of 65 | 2.3 M older adults, Ontario |
| Largest single dyad — oral iron followed by a new laxative | 11.9% | 95% CI 11.7–12.1 |
| The amlodipine case above — calcium channel blocker followed by a new diuretic | 7.5% | 95% CI 7.3–7.6 |
| Prevalence of an active cascade, community-dwelling | 2.1% | Ireland, 9 dyads |
| Possible cascade, multimorbid hospital inpatients | 16.3% | Prospective, 2025 |
| Excess emergency visits and hospitalisations for patients in a cascade | HR 1.21 | 95% CI 1.02–1.43 |
| Prevalence of all 65 dyads in a US community population | Unknown | No study has been done |
Rates rise with illness burden: roughly one in forty among community-dwelling older adults, roughly one in seven among sick, heavily medicated inpatients. The last row is the gap we exist to close. Figures circulating elsewhere — that a quarter of older adults are in an active cascade, or that cascades account for hundreds of billions of dollars — do not survive a check against their stated sources, and we do not use them.
How it works
Five conditions, tested against a dated medication and problem record.
The input is what a health plan or health system already holds: dated medication fills, coded problems and laboratory values. A patient is flagged only when all five conditions hold.
-
Culprit drug present
The drug capable of causing the effect, resolved from brand names to ingredient level.
-
Treatment drug started later
The second drug appears after the first, inside the window defined for that dyad.
-
Culprit still active
If the first drug has already stopped, the cascade is broken and there is nothing to flag.
-
Exclusions cleared
An independent indication for the second drug rules the case out. This is where false positives are prevented.
-
Effect coded between them
The adverse effect appears in the record between the two prescriptions.
Rules are data, not code
All 65 cascades are stored as structured rules — drug sets, adverse-effect definitions, temporal windows and exclusion logic, 37 fields each. A clinical reviewer can change a window or add an exclusion without an engineering release. That is deliberate: the parameters are clinical judgments and they should be editable by clinicians.
The hard part is attribution
Detecting that one drug preceded an effect preceded another drug is deterministic and straightforward. Deciding whether the second drug was prescribed because of the first, rather than for a genuine independent reason, is not — and it is what sets the false-positive rate. Alerts fail through clinician override, not through poor detection. That question is the centre of our research programme.
Where we are
What is built, and what is not true yet.
Clinical software earns trust by being precise about its own limits. So, plainly:
Built and running
- The detection engine, covering all 65 consensus cascades, in two forms: a Python package for batch files and a self-contained browser build.
- The full rule specification — 65 cascades, 37 fields each, complete and ready for independent clinical adjudication.
- Terminology resolution across 274 brand and ingredient names, with ICD-10 matching for adverse effects.
- An automated test suite, including near-miss cases — patients who meet every structural condition but are correctly prescribed.
Not yet
- No independent clinical validation. The rule parameters are our own clinical judgment, drawn from the consensus paper. No outside panel has adjudicated them.
- No real patient data. Everything run to date is synthetic.
- No deployed integration. Nothing is connected to any electronic health record.
- No design partner signed, and no published accuracy figure. Both are the work of the next twelve months.
Demonstration
Ten constructed cases, including ones that should not fire.
The demonstration below runs entirely in your browser on fictional patients. It includes near-miss cases on purpose: patients who satisfy every structural condition of a cascade but whose second drug is independently indicated. Those are the cases that decide whether a tool like this is usable, and a demonstration that only shows the easy ones is not worth much.
Open the demonstration Demonstration only. Not a medical device, and not for clinical use.Evidence
Everything above traces to three papers.
- The 65-cascade expert consensus Rochon PA, et al. Potentially inappropriate prescribing cascades. European Geriatric Medicine 2025;16:1573–1584. Modified Delphi consensus, endorsed by the European Geriatric Medicine Society. The source of the target set.
- Population measurement of all 65 Rochon PA, et al. BMJ 2026;394:e100499. All 65 dyads measured across 2,297,942 older adults; 49 exceeded 1% annual incidence and 24 were prioritised by population burden. Read it
- Harm attributable to a cascade Rochon PA, et al. Journal of the American Geriatrics Society 2024;72:467–478. Propensity-matched analysis finding a hazard ratio of 1.21 (95% CI 1.02–1.43) for emergency department visits and hospitalisations.
Community prevalence figures are from Doherty AS, et al., in the Irish Longitudinal Study on Ageing and in Dutch general practice; inpatient figures from the REPOSI registry and a 2025 prospective prevalence study. We are not affiliated with any of these research groups, and citing their work does not imply their endorsement.
Who we are
A pharmacist-led company, because the hard input here is clinical.
Deciding how many days may separate two prescriptions before the link stops being plausible, and which independent indications should rule a case out, is pharmacist work. It is the part of this problem that cannot be solved by engineering alone.
Working on this problem?
We are looking for clinical collaborators for the adjudication study, and for health plans and health systems willing to be design partners on de-identified data. If that is you, or if you want to pick holes in the rule specification, we would like to hear from you.
Senetika, Inc. · Delaware C-corporation · St. Louis metro and Metro East Illinois