Early warning.
No action
required.
Tenor monitors CHF patients passively, through their Apple Watch and AirPods, and detects physiological deterioration days before a crisis. The patient lives their life. Tenor watches.
CHF kills because it hides.
26 million people live with congestive heart failure right now, and half of them will end up in the hospital within a year of their diagnosis. We think most of that is preventable, if the drift is caught early enough to act on. Every early-warning tool out there still asks something of the patient: log this, wear that, remember to check in. But the people who need monitoring most are often the least able to keep up a routine. So we built Tenor to ask for nothing at all.
with CHF
within one year
before acute event
attributable to CHF
Five layers.
One inference.
Tenor blends what it hears with what it measures: acoustic biomarkers alongside physiological signals, run through a model built around you specifically. We never let one signal make the call by itself. The Multi-Modal Deterioration Index looks for patterns that show up together, small changes that wouldn't mean much on their own, but add up to something worth paying attention to.
Passively captures ambient speech during natural conversation. Voice Activity Detection with speaker verification isolates the patient's voice. Raw audio is discarded immediately; only features are retained, on-device.
HRV, SpO₂, respiratory rate, ECG single-lead, and wrist temperature provide continuous physiological telemetry. All sensors are already present on commercially available hardware, so no new devices are required.
Declining daily step count and sleep fragmentation from orthopnea are recognized CHF prodromal signs. Patients slow down days before other symptoms appear, a pattern invisible to them but clear to continuous monitoring.
All processing runs locally. Modality-specific encoders feed a cross-modal attention mechanism producing the MMDI score. The personalized model continuously calibrates, learning your specific deterioration signature over weeks.
Notifications reach patient, family, and clinical team through contextually appropriate channels. The system closes the entire loop: alert → charts transmitted → appointment scheduled → transport arranged → physician briefed. No manual steps.
Confirmed deterioration events are retained at elevated fidelity as a per-patient episode library. Future alerts compare against this record for similarity-based confidence adjustment, making the system more precise over time, not less.
vs. personalized baseline
Co-occurring subthreshold changes across modalities produce a suprathreshold alert. A 1.8σ voice deviation, a 12ms HRV decline, and a 2% SpO₂ drift would each pass unnoticed on their own. When they co-occur across the same 72-hour window, MMDI amplifies their convergence into a high-confidence signal. This cross-modal attention mechanism sets Tenor apart from single-modality approaches, and it sits at the core of Softsign's patent claims.
Tuesday morning. Walter is making coffee. His Apple Watch has been tracking his sleep since 3am, when his respiratory rate began climbing. His AirPods, worn during his morning walk, captured 40 minutes of ambient audio. His voice has been subtly different for three days. Not to his wife. To Tenor.
"Walter, I've been noticing some changes over the past few days that I'd like your doctor to look at. Would you like me to send your charts to Dr. Patel?"
Walter says yes. Tenor sends a structured summary to his cardiologist's clinical system. She adjusts his diuretic remotely. Walter never goes to the emergency room. He never feels the breathlessness. He has no idea how close he came.
Walter's daughter is at work. She gets a quiet notification, not an alarm, not a scare.
"Your father's health indicators have been shifting over the past few days. His care team has been notified. I'll keep you updated. No action needed from you right now."
She breathes. She texts her father. She doesn't have to get on a plane. She doesn't spend the next three days paralyzed with worry. She is informed, not alarmed. That distinction matters enormously, and it's one that few tools deliver today.
One patient.
One family.
Every decision.
To build Tenor right, we keep one person in mind. Not a demographic. Not a persona. A real human being, and everyone who loves them.
The things we refuse
to become.
We're writing these down now, at the very start, so that later, when it would be easier not to, we still hold ourselves to them.
Building in the open.
Work done.
Phase 0 is done. We've built the infrastructure to validate Tenor's core signal, and we've already tested it with real participant data.
SwiftData + HealthKit app deployed on iPhone and Apple Watch Ultra. All seven modalities collected and deduplicated. Validated with real participant data.
Browser-based admin dashboard. No server. Drag-and-drop import, synchronized scrub-line across all sensors, zoom, and localStorage persistence.
Full provisional application covering six inventive concepts. Nine USPTO-style technical drawings. Priority date filing imminent following attorney review.
20–40 CHF patients. Observational. Existing Apple hardware, so no device procurement is needed. Clinical partner: Jon, Chief Innovation Officer, University of Iowa Hospitals.
Five phases from
concept to platform.
CHF is our proof of concept, not our ceiling. The architecture behind Tenor (continuous passive monitoring, a model that adapts to you, care coordination that happens quietly in the background) is how we think AI should show up in health care, for every chronic condition, not just this one.
You shouldn't have to
fight this on your own.
The sensors already exist. So does the silicon, the privacy framework, the whole ecosystem behind it. What was missing was someone willing to put those pieces together for something that actually matters to real people. We decided that's us, starting now.