Every model. Open card. Open metrics. Open limitations.
Most healthcare AI is a black box that says “trust us.” Materna Health Solutions publishes a model card for every algorithm in production. Inputs, outputs, validation set, fairness metrics, refresh schedule, and what each model is not for.
5
Models in design
pending
Validation AUC
Reported once measured
Quarterly
Fairness audit cadence
From go-live onward
0
Patients in validation
Pre-launch, no lived data yet
Principles
The rules every Materna algorithm has to follow.
Decision support, never decision-making.
No Materna model writes orders, schedules tests, or sends messages without a clinician in the loop. An algorithm can flag, suggest, prioritize. It never decides.
Plain-language explanations, on demand.
Any patient or clinician can ask "why" on any flag and get the contributing factors in plain language, in EN or ES. The model card is one click away.
Bias measurement is non-optional.
Every model is evaluated quarterly for equalized odds across language, race, and rural/urban subgroups. Gaps over 3 percentage points trigger a hold.
Safety pathways never run through the score.
Self-harm, severe bleeding, decreased fetal movement, and acute symptoms always go to the safety pathway. The score does not gate anything urgent, ever.
Validation, refresh, and rollback.
Every model has a documented validation set, refresh schedule, and rollback procedure. If a model degrades, we revert. We do not paper over it.
Open documentation.
Every model card is on this page. Every change is logged. Researchers can request the methodology document and our published validation studies.
Model cards
One card per model. Five models. No skeletons in the closet.
Preeclampsia early-warning
Predicts probability of developing preeclampsia in the next 4 weeks based on serial blood pressure, urine protein, weight gain trend, and structured history.
Inputs
- Last 30 days of blood pressure readings
- Last 30 days of weight, against trimester-adjusted curve
- Most recent urine protein
- Maternal age, BMI, parity
- Prior preeclampsia, chronic HTN, diabetes, autoimmune disease
Output
Risk band: low / moderate / high. With confidence interval and the top 3 contributing factors.
Intended use
Decision support for the OB/MFM. Surfaces patients who warrant closer monitoring or early aspirin if not already on it.
Not intended for
Diagnosis. Not a substitute for clinical judgment. Not for triage of acute symptoms (severe headache, vision changes, RUQ pain).
Validation
Cohort
In development. Validation will be done on a border-corridor cohort with public methodology.
N
pending
Fairness
Language gap
Will be published once measured (EN vs. ES)
Race / ethnicity gap
Will be published once measured. Hispanic vs. non-Hispanic white is the primary stratification.
Method
Equalized odds, evaluated quarterly. Disparate-impact ratio will be published with each refresh.
Guardrails
- Never auto-orders medication. Always clinician-in-loop.
- Never suppresses safety alerts based on score.
- Provider can override with reason captured in chart.
- Patient can request explanation in plain language EN + ES.
Preterm-birth risk
Estimates probability of preterm birth (< 37 weeks) using cervical length, prior preterm history, modifiable risk factors, and structured social determinants.
Inputs
- Cervical length on TVUS at 18-24 weeks
- Prior preterm birth or second-trimester loss
- Smoking, body mass index, infections
- Distance to nearest delivering hospital
- Housing stability, food access, transportation
Output
Risk band, with eligibility flag for 17-OHP, vaginal progesterone, or cerclage.
Intended use
Decision support to prompt MFM consult and consider evidence-based prophylaxis when indicated.
Not intended for
Confirming labor or rupture of membranes. Acute symptoms route to labor floor, never the algorithm.
Validation
Cohort
In development. Plan: tertiary referral plus FQHC border-corridor cohort, sensitivity-prioritized.
N
pending
Fairness
Language gap
Will be published once measured
Race / ethnicity gap
Will be published once measured. Hispanic vs. non-Hispanic white primary; Black vs. white tracked.
Method
Equalized odds with sensitivity-prioritized threshold for high-risk subgroups.
Guardrails
- Top contributing SDOH factors trigger care-coordinator referral, not just an alert.
- Distance-to-hospital factor surfaces a transportation benefit eligibility check.
- Provider review required before any prophylaxis order.
Gestational diabetes risk
Estimates probability of GDM at 24-28 weeks, used for early counseling and screening compliance.
Inputs
- Pre-pregnancy BMI, age, parity
- Family history of T2DM
- Prior GDM or large-for-gestational-age baby
- PCOS, racial/ethnic factors per ACOG
- First-trimester fasting glucose if available
Output
Risk band. Triggers earlier OGTT (16-20 weeks) when high.
Intended use
Identify candidates for early GDM screening per ACOG guidance and initiate nutrition referral.
Not intended for
Replacing OGTT. Always followed by gold-standard diagnostic test.
Validation
Cohort
In development. Plan: ACOG-aligned cohort with stratification by ethnicity given known disparate prevalence.
N
pending
Fairness
Language gap
Will be published once measured
Race / ethnicity gap
Will be published once measured. Particular care taken given known disparate prevalence.
Method
Equalized odds.
Guardrails
- Never used to deny screening. Used only to prioritize earlier testing.
- Patient sees the explanation in plain language.
Postpartum depression rising-risk
Combines screen scores, symptom narratives, sleep, weight trajectory, and follow-up adherence to flag patients trending toward PPD before the formal EPDS catches it.
Inputs
- Most recent PHQ-9, EPDS, GAD-7
- Sleep totals (last 14 days)
- Weight trajectory and feeding pattern
- No-show or canceled visits in postpartum window
- Voice-companion narrative tone (with consent)
Output
Soft flag with rationale. Routes to coordinator for outreach.
Intended use
Trigger an outreach call before the patient stops responding. Assist clinician judgment.
Not intended for
Clinical diagnosis. Self-harm phrases always escalate independently to safety pathway.
Validation
Cohort
In development. Plan: postpartum cohort with EN and ES language stratification, voice-tone features opt-in only.
N
pending
Fairness
Language gap
Will be published once measured (EN and ES will be validated separately).
Race / ethnicity gap
Will be published once measured.
Method
Equalized odds. Voice tone features will be stratified by primary language.
Guardrails
- Voice features only used with explicit consent. Always opt-in.
- Self-harm escalates regardless of overall score.
- Soft flag never appears to anyone outside the care team.
Fetal growth restriction watch
Combines serial ultrasound biometry, Doppler, fundal height trend, and maternal factors to flag growth concerns.
Inputs
- Last 3 ultrasound EFW measurements
- Umbilical artery Doppler
- Maternal HTN, smoking, autoimmune
- Placental position and prior history
Output
Risk band, with MFM referral suggestion for moderate or higher.
Intended use
Surface the patient who needs a growth check or MFM consult, before the next routine visit.
Not intended for
Replacing ultrasound interpretation. Always read by a qualified sonographer or MFM.
Validation
Cohort
In development. Plan: tertiary + community OB cohort with growth-velocity stratification.
N
pending
Fairness
Language gap
Will be published once measured
Race / ethnicity gap
Will be published once measured
Method
Equalized odds.
Guardrails
- Never autonomously orders imaging. Always clinician confirms.
- Output never visible to the patient until provider has reviewed.
Adverse-event reporting
We log every override, every miss, every save.
Every clinical interaction with a Materna algorithm is logged. Overrides are captured with reason. Confirmed false negatives and false positives feed back into the next quarterly retrain. We publish an annual safety report.
Always-on safety pathways
These never run through any algorithm. They route directly to the on-call line, regardless of any model output:
- Self-harm or suicidal ideation
- Severe headache, vision changes, RUQ pain (preeclampsia signs)
- Heavy bleeding
- Decreased fetal movement (after 28 weeks)
- Domestic violence keywords