Technical Architecture

This document describes SmartShot's technical design with honest assessments of what exists, what's emerging, and what requires breakthroughs nobody has made yet. Every claim references published research or commercially available technology. Where gaps exist, they are named.

1. System Architecture

SmartShot is a four-layer system. Each layer has different maturity. Some layers use proven technology. Others do not exist yet.

Layer 1: Sensor Array

The sensor layer must continuously monitor physiology and detect the onset of three distinct emergencies: anaphylaxis, opioid-induced respiratory depression, and acute myocardial infarction. Each condition requires different sensing modalities.

Cardiac Monitoring (ECG + PPG)

Single-lead ECG patches exist today from companies like LifeSignals, VitalConnect, and Masimo. The LifeSignals 2A Biosensor provides continuous 2-channel ECG, SpO2, pulse rate, respiration rate, and body temperature in a clinical-grade wearable form factor.

For ST-segment elevation detection — the hallmark of acute myocardial infarction — single-lead wearable ECG shows limited sensitivity: 34% with high specificity (100%) in one study. Multichannel acquisition dramatically improves this, with sensitivity of 93-94% and specificity of 92-100%, but requires manual repositioning of the device across the chest. A body-worn device in a fixed position would be constrained to single-lead performance unless multiple electrode patches were used.

Photoplethysmography (PPG) sensors achieve SpO2 accuracy within RMSE 2.1%, meeting ISO 80601-2-61 clinical standards. Heart rate accuracy reaches 95.47 +/- 4.31% in current wearables.

Respiratory Monitoring

Accelerometer-based respiratory monitoring is the most proven approach for wearables. The University of Washington's closed-loop naloxone injector prototype used paired body accelerometers to track respiration and detect opioid-induced apnea. In controlled testing at a certified injection facility, the sensors accurately tracked respiration rates and detected non-medical, opioid-induced apnea events.

Strain-gauge chest bands provide higher-fidelity respiratory data but add bulk and complexity. Impedance pneumography offers another path but requires chest-mounted electrodes.

For SmartShot, respiratory rate monitoring must detect cessation of breathing within 15-30 seconds to allow time for naloxone injection before irreversible hypoxia. Current accelerometer-based systems can achieve this window.

Anaphylaxis Detection

This is the hardest sensing problem. No laboratory test can diagnose anaphylaxis in real time. The two primary biomarkers — histamine and tryptase — have significant limitations:

Histamine: pooled sensitivity of 0.76, specificity of 0.69. Detectable within 15 minutes of onset but has a short biological half-life.

Tryptase: pooled sensitivity of only 0.49, specificity of 0.82. Longer half-life (2 hours) but too insensitive for stand-alone use.

Combined histamine + tryptase detection achieves 0.93 sensitivity, but requires blood sampling — not currently possible in a non-invasive wearable.

The Wyss Institute's Project Abbie (Harvard/Boston Children's Hospital) developed "abbieSense," a sensor that detects therapeutically relevant histamine levels and determines reaction severity within five minutes. The device would also monitor skin response, temperature, and cardiac rhythms. This is the closest existing prototype to SmartShot's anaphylaxis detection needs, but it remains pre-clinical and no 2025-2026 status updates are publicly available.

Surrogate detection through skin conductance changes, heart rate variability, and respiratory patterns could provide earlier warning, but with lower specificity. The fundamental challenge: anaphylaxis shares physiological signatures with anxiety, exercise, and heat exposure.

Layer 2: Decision Engine

The decision layer runs on-device ML models that analyze streaming sensor data and determine when to trigger medication delivery. This is where the system either saves lives or kills people. There is no middle ground.

Algorithm Requirements

The decision engine must achieve near-zero false negatives (missing a real emergency is fatal) while maintaining a false positive rate low enough that autonomous injection is safe. These goals are in direct tension.

Published performance benchmarks for wearable anomaly detection:

Arrhythmia detection via PPG: 93% true positive rate, 54% true negative rate. The 46% false positive rate is unacceptable for autonomous drug injection.

Real-time CNN-based arrhythmia detection: achievable on embedded hardware with 210 KB memory footprint and 94.8ms inference time. Accuracy of 86.1%.

Hospital alarm systems: 27.4% false alarm rate overall, rising to 91.4% for acute life-threatening alarms. This is the fundamental problem — systems tuned to catch every real event generate enormous noise.

HRV-based anomaly detection: 85.7% sensitivity with false positive rate of 0.62/hour. At that rate, a device worn 24 hours would generate roughly 15 false positives per day.

Required Thresholds

For SmartShot to be viable, the decision engine must achieve:

Sensitivity (true positive rate): >99.5% — missing 1 in 200 real events is already too many, but perfection is not achievable.

Specificity (true negative rate): >99.99% — at 99.9%, a device monitoring continuously would generate roughly one false injection per 42 days. At 99.99%, roughly one per 14 months. Neither is acceptable for autonomous epinephrine injection, but the latter enters a tolerable range if the control hub adds a human confirmation step.

No published wearable system achieves these thresholds today. Most are two to three orders of magnitude away from the required specificity.

Multi-Signal Fusion

The path to adequate accuracy is combining multiple sensor streams. A cardiac event that triggers ECG anomaly + SpO2 drop + respiratory rate change simultaneously is far more likely to be real than any single-signal alarm. Sensor fusion algorithms using ensemble methods and temporal correlation can dramatically reduce false positives compared to single-channel detection.

This is mathematically sound. The engineering challenge is validating the combined model against sufficient real-world data, across diverse patient populations, in ambulatory conditions with motion artifacts.

Edge Processing

All decision logic must run on-device. A cardiac arrest patient cannot wait for a cloud API round-trip. Latency budget from detection to injection initiation: under 500 milliseconds for cardiac events, under 30 seconds for respiratory depression, under 5 minutes for anaphylaxis.

Current edge AI chips can run inference in under 100ms with power consumption in the microwatt range. BrainChip's neuromorphic processors and similar platforms support continuous monitoring on minimal power. The compute is not the bottleneck. The training data and model validation are.

Layer 3: Automatic Medication Injection (AMI)

The AMI subsystem is a miniaturized auto-injector that delivers pre-loaded medication subcutaneously on command from the decision engine or control hub.

Injection Mechanism

Four mechanism types dominate the wearable injector patent landscape (163,000+ US patents filed as of 2025):

Spring-based: a compressed spring drives a plunger to push medication through a needle. Simple, reliable, no battery required for the injection stroke itself. This is the mechanism used in EpiPen and similar devices. Disadvantage: single-use. Once the spring fires, the cartridge must be replaced.

Motor-driven: a small DC motor or stepper motor drives the plunger. Allows variable delivery rates and multiple injections from a single reservoir. Requires battery power for injection. More complex, more failure modes.

Expanding battery: uses a gas-generating electrochemical cell to create pressure. Novel approach but limited commercial validation.

Rotary pump: peristaltic or gear pump delivers medication continuously or in boluses. Most flexible but largest form factor.

For SmartShot, the motor-driven approach is the strongest candidate. It allows multiple injection doses from a single cartridge (critical for naloxone, where repeat dosing is common) and variable delivery rates. The UW prototype opioid overdose device used a commercial injector adapted for automated triggering.

Drug Cartridge Design

Each SmartShot variant must carry a specific medication:

Epinephrine (anaphylaxis): standard adult dose is 0.3 mg in 0.3 mL. Volume is small. The challenge is stability, not volume.

Naloxone (opioid overdose): standard dose is 0.4 mg, but higher doses (2-4 mg) are increasingly used for fentanyl analogs. May require multiple injections. Total volume: 1-2 mL for multi-dose capability.

Nitroglycerin (cardiac): sublingual is the standard route. Subcutaneous nitroglycerin delivery is not an established clinical pathway. This is a significant pharmacological gap — SmartShot's cardiac variant may need to deliver a different medication (e.g., aspirin is not injectable subcutaneously either) or use a different delivery route entirely.

Sterility

The injection site and needle must remain sterile for weeks or months while worn on the body. This requires a sealed needle deployment system with a sterile barrier that breaks only at the moment of injection. Similar engineering exists in insulin pump infusion sets, which maintain sterility for 3-day wear periods. Extending this to 30+ days is an unsolved engineering problem.

Needle Deployment

The needle must penetrate skin reliably across varying body compositions. Subcutaneous injection depth ranges from 5-15mm depending on patient body fat percentage. Spring-loaded needle deployment with pre-set depth is the standard approach. Retractable needles that withdraw after injection reduce infection risk and accidental needlestick.

Layer 4: Communication

The communication layer connects the device to the control hub and emergency services. Three tiers of connectivity, each with different latency and reliability profiles.

Bluetooth Low Energy (BLE)

Primary link between device and patient's smartphone. BLE 5.0 provides 2 Mbps throughput, 240m range in open air (reduced to 10-30m through walls), and low power consumption. Latency: 7.5ms connection interval. Sufficient for streaming vital sign data and receiving override commands.

Limitation: requires the patient to carry a paired smartphone. In an overdose scenario, the patient's phone may not be present or charged.

Cellular (LTE-M / NB-IoT)

Direct device-to-tower communication for when BLE link is unavailable. LTE-M provides 1 Mbps with latency of 10-15ms. NB-IoT offers lower power but higher latency (1.6-10 seconds). Both require embedded SIM and antenna, adding roughly 2-3 cm3 to device volume and 50-100 mW to power budget during transmission.

Cellular is the critical link for control hub communication. Without it, the device operates autonomously — which means the human oversight layer is bypassed.

Satellite (LEO)

For rural and wilderness coverage where cellular is unavailable. Emerging LEO satellite IoT networks (e.g., Globalstar, Swarm) offer 1-10 minute message latency. Too slow for real-time override but sufficient for post-event notification and location reporting.

2. The Control Hub

The control hub is SmartShot's critical differentiator. Every other layer is an engineering problem. The control hub is an operations problem — and the reason SmartShot can exist within current regulatory and liability frameworks while a fully autonomous device cannot.

Function

The control hub is a 24/7 staffed medical command center. Licensed paramedics and nurses monitor incoming alerts from SmartShot devices and make the final injection decision for cases where the device's confidence level falls below the autonomous threshold.

This creates a tiered response:

Tier 1 — Autonomous: Device confidence >99.9%. Patient is apneic for >60 seconds with falling SpO2. The device injects naloxone immediately and notifies the hub. No human delay.

Tier 2 — Confirmed: Device confidence 95-99.9%. Hub receives alert, reviews streaming vitals, confirms injection within 15-30 seconds. Operator can override.

Tier 3 — Escalated: Device confidence 80-95%. Hub contacts patient (audio through device or phone), assesses situation, decides whether to inject or dispatch EMS.

Tier 4 — Monitoring: Device confidence <80%. Anomaly logged. Patient not disturbed. Data reviewed in batch for algorithm improvement.

Staffing Model

Staffing scales with device population. Assuming average alert rates:

Tier 1 events: estimated 0.001% of devices per day (1 per 100,000 device-days). At 10,000 active devices, one event every 10 days.

Tier 2 events: estimated 0.01% per day. At 10,000 devices, roughly one per day.

Tier 3 events: estimated 0.1% per day. At 10,000 devices, roughly 10 per day.

Tier 4 events: estimated 1-5% per day. Handled by automated logging. No real-time staff required.

Initial staffing: 2 operators 24/7 (4.2 FTE with coverage for breaks and shift overlap). Scale trigger: add 1 FTE per 5,000 active devices.

Infrastructure

• Redundant internet connections (minimum two ISPs, different last-mile technologies)
• Backup power (UPS + generator, 72-hour fuel capacity)
• HIPAA-compliant workstations with real-time vital sign dashboards
• Direct integration with 911/PSAP systems for EMS dispatch
• Recorded audio capability for patient communication
• Geographic redundancy: two physical sites, active-active configuration

Regulatory Position

The control hub transforms SmartShot from a Class III autonomous drug delivery device into something closer to a telemedicine-supervised drug administration system. This distinction matters for FDA pathway. A device that always has a licensed clinician in the loop — even if the loop is sometimes bypassed for the most critical cases — occupies different regulatory ground than a fully autonomous injector.

This is not regulatory arbitrage. It is a genuine safety layer. Humans catch edge cases that algorithms miss.

3. Feasibility Summary

EXISTS TODAY:

• Single-lead continuous ECG monitoring
• PPG-based SpO2 and heart rate
• Accelerometer-based respiratory rate tracking
• On-device ML inference (edge AI chips)
• Single-dose auto-injection mechanisms (spring-based)
• BLE and LTE-M communication modules
• 24/7 medical monitoring center operations
• Subcutaneous epinephrine and naloxone formulations

IN DEVELOPMENT (2-5 years):

• Multi-signal fusion algorithms with clinically validated accuracy
• Motor-driven multi-dose wearable injection systems
• Extended-duration sterile needle containment (>3 days)
• Sub-30-second human-in-loop confirmation workflow
• Tiered autonomous/supervised injection decision framework
• Neuromorphic ultra-low-power continuous sensing

REQUIRES BREAKTHROUGH:

• Non-invasive real-time anaphylaxis biomarker detection
• 99.99% specificity for autonomous injection decisions
• Subcutaneous cardiac emergency medication delivery
• Combined sensor array + drug reservoir + injection + battery + comms in a body-worn form factor under 100g
• Regulatory framework for conditionally autonomous drug delivery

4. Engineering Challenges

These are the problems that determine whether SmartShot is built in five years or fifteen.

Battery

Continuous sensing at 100-200 Hz sampling rates consumes significant power. Current medical wearables use 100-400 mAh batteries and achieve 1-7 day battery life for sensing alone. SmartShot adds three power demands that standard wearables do not have:

1. Cellular radio: 50-100 mW during transmission bursts. Even with duty cycling, this adds 20-30% to daily power consumption.

2. Injection mechanism: motor-driven injection requires a brief but intense current draw — estimated 500-1000 mA for 2-5 seconds. A 300 mAh battery can support this, but only if it is not already depleted by sensing.

3. Always-on ML inference: current edge AI chips consume 0.5-5 microwatts in monitoring mode with wake-up in 50-200 microseconds. This is manageable.

Target battery life: 7 days minimum between charges. This is achievable for the sensing and communication functions with a 400 mAh battery. The injection reserve must be maintained separately — either a dedicated capacitor bank or a mechanically isolated battery cell that cannot be drained by routine operations. If the device runs out of charge for sensing, the injection mechanism must still function.

Miniaturization

SmartShot must fit on the body without impeding daily life. The target form factor is roughly the size of a continuous glucose monitor: 35mm diameter, 10mm thick. Current CGMs contain only a sensor and BLE radio. SmartShot must fit:

• Multi-sensor array (ECG electrodes, PPG sensor, accelerometer, temperature)
• Edge AI processor
• BLE + cellular radio with antenna
• Drug reservoir (0.3-2 mL depending on medication)
• Injection mechanism (motor, plunger, needle assembly)
• Battery (400+ mAh)
• Sterile barrier and needle containment

This will not fit in a CGM form factor. A more realistic target is the size of an insulin pump: roughly 80 x 50 x 15mm, weighing 80-120g. This is wearable but not invisible. Patients would clip it to a waistband or wear it in an adhesive body patch.

Achieving the CGM form factor would require advances in battery energy density, motor miniaturization, and component integration that do not exist on any current development roadmap.

Drug Stability

Body-worn devices expose medication to skin-surface temperatures (32-37 degrees C), motion, and ambient light. Published stability data:

Epinephrine: stable at temperatures up to 28.9 degrees C for 45 days. Tolerates temperature spikes to 51.7 degrees C for cumulative 795 minutes. However, at sustained 50 degrees C, rapid degradation occurs. Standard shelf life is 18 months at controlled room temperature (20-25 degrees C). Body-worn exposure at 32-37 degrees C would likely reduce effective shelf life to 30-90 days. The cartridge replacement cycle would need to match.

Naloxone: more durable. Ampoules show no concentration changes after 28 days of heat exposure or freeze-thaw cycling compared to room temperature controls. Body-worn stability is likely adequate for 90+ day cartridge life.

Nitroglycerin: extremely unstable. Sublingual tablets expire 3 months after opening. The molecule is volatile and light-sensitive. Body-worn delivery of nitroglycerin is not feasible with current formulations. This is one reason the cardiac variant of SmartShot may need to use a different pharmacological approach entirely.

Biocompatibility

Extended skin contact raises dermatitis risk. Medical-grade adhesives used in CGMs and insulin pump infusion sets cause skin reactions in 10-20% of users over multi-day wear. SmartShot's larger footprint increases contact area and reaction likelihood.

Injection site rotation is standard practice for insulin pumps, where patients change infusion sites every 3 days. SmartShot would need to either rotate injection sites (requiring patient repositioning of the device) or engineer a mechanism that can inject at slightly different locations within its footprint. The latter adds mechanical complexity.

Biocompatible housing materials (medical-grade silicone, polycarbonate) are well established. The novel challenge is the sterile needle barrier — it must remain sealed during weeks of body contact, sweat exposure, and mechanical stress, then break cleanly at injection time.

False Positive Mitigation

This is the defining engineering challenge. What accuracy is required before you let a machine inject medication into a human body without asking first?

The answer depends on the medication:

Naloxone: relatively safe. Side effects in a non-overdosing patient include nausea, agitation, and brief withdrawal symptoms in opioid-dependent users. A false positive injection is unpleasant but not dangerous. This makes naloxone the strongest candidate for autonomous injection with lower confidence thresholds.

Epinephrine: moderate risk. False positive injection causes tachycardia, hypertension, anxiety, tremor. In patients with cardiac disease, unnecessary epinephrine can trigger arrhythmia. False positive tolerance is lower.

Nitroglycerin: significant risk if delivered systemically (sublingual is self-limiting). Hypotension, syncope, falls. In combination with certain medications (PDE5 inhibitors), potentially fatal.

The risk asymmetry suggests a phased approach: launch SmartShot Naloxone first (highest false positive tolerance, most proven detection algorithm, most established injection pharmacology), then SmartShot Epi (moderate difficulty), then cardiac variant (hardest on every dimension).

5. Security and Privacy

Threat Model

SmartShot is a device that can inject medication into a human body based on electronic commands. If compromised, it is a weapon. This is not hyperbole. The threat model must assume a motivated, technically sophisticated adversary.

Attack surfaces:

1. BLE radio: attacker within 30m could attempt to spoof injection commands. Mitigation: mutual authentication, encrypted command channel, injection commands require cryptographic signature from either the on-device decision engine or the authenticated control hub. BLE pairing with out-of-band confirmation.

2. Cellular channel: man-in-the-middle between device and control hub. Mitigation: TLS 1.3 with certificate pinning. Device authenticates hub; hub authenticates device. No fallback to unencrypted communication.

3. Physical tampering: attacker with physical access to device could modify firmware or drug cartridge. Mitigation: tamper-evident seals, hardware security module (HSM) for firmware verification, cartridge authentication via RFID or cryptographic chip.

4. Cloud infrastructure: if control hub servers are compromised, an attacker could issue injection commands to any connected device. Mitigation: injection commands are cryptographically signed with keys stored in the device's HSM. Even a compromised server cannot forge valid injection commands without the device-specific key.

5. Supply chain: compromised firmware during manufacturing. Mitigation: secure boot chain, manufacturer attestation, independent security audit of firmware before deployment.

FDA Cybersecurity Requirements

The FDA's final guidance on medical device cybersecurity (June 2025) requires:

• Alignment with NIST cryptographic guidelines and FIPS 140-3
• Defense-in-depth architecture: secure defaults, least-privilege access, multi-factor authentication, strong encryption, segmentation of critical functions, anomaly detection telemetry
• Software Bill of Materials (SBOM) for all device components
• Documented plans for regular and emergency software updates
• Ongoing risk assessment and vulnerability tracking
• Compliance with ISO 13485:2016 under QMSR by February 2026

SmartShot's architecture must meet all of these from day one. Retrofitting security is not an option for a life-safety device.

Edge vs. Cloud

The decision engine runs on-device. This is not negotiable. A cardiac arrest patient has minutes. A cloud API call over cellular has latency of 50-200ms under ideal conditions, but cellular connectivity is not guaranteed.

Architecture: all life-critical decision logic runs on the device's edge processor. The cloud connection is used for:

• Streaming vitals to the control hub (monitoring, not decision-making)
• Receiving human override commands (additive safety, not required for function)
• Firmware updates (asynchronous, verified before installation)
• Post-event data upload (forensic and quality improvement)

If the device loses all connectivity, it continues to function autonomously at Tier 1 thresholds. The control hub adds safety margin. It is not a dependency.

Privacy

SmartShot continuously records physiological data. This is HIPAA-protected health information. Storage and transmission must comply with the HIPAA Security Rule.

On-device: data encrypted at rest using AES-256. Rolling 72-hour buffer of raw sensor data. Older data transmitted to cloud and purged from device.

In transit: TLS 1.3 for all transmissions.

At rest (cloud): encrypted storage, access controls, audit logging. Patient data segregated by device. De-identified data used for algorithm improvement only with explicit patient consent.

6. R&D Roadmap

Six phases. Honest timelines. Each phase has a gate that must be passed before proceeding. Failure at any gate means the program pauses until the problem is solved.

Phase 1: Sensor Validation (Months 1-18)

Objective: prove that a wearable sensor array can detect each target condition with sufficient accuracy to justify continued investment.

Work:

• Select and benchmark commercial ECG, PPG, accelerometer, and temperature sensors
• Build multi-sensor prototype boards (not wearable — bench-top validation)
• Collect 10,000+ hours of labeled physiological data from at-risk patient populations
• Test detection algorithms against labeled data
• Publish false positive and false negative rates for each condition

Gate: cardiac event detection sensitivity >95% and specificity >99% in bench-top testing. Respiratory depression detection sensitivity >98% with <1 false positive per 8 hours. Anaphylaxis detection: determine whether non-invasive biomarker sensing is viable or whether surrogate detection is the only path.

Cost estimate: $2-4M
Key risk: anaphylaxis detection may prove infeasible without invasive biomarker sampling

Phase 2: Decision Algorithm Development (Months 12-30, overlapping Phase 1)

Objective: develop and validate ML models for autonomous injection decisions.

Work:

• Train multi-signal fusion models on Phase 1 data
• Develop confidence scoring system aligned with control hub tier structure
• Test against holdout datasets and synthetic adversarial scenarios
• Simulate false positive injection rates at scale (10,000 device-years)
• Engage FDA pre-submission meeting to discuss algorithm validation requirements

Gate: simulated false positive injection rate <1 per device-year for naloxone variant. Sensitivity >99% for respiratory depression.

Cost estimate: $3-5M
Key risk: insufficient training data diversity. Algorithms trained on clinical populations may fail in real-world ambulatory conditions.

Phase 3: AMI Prototype (Months 18-36)

Objective: build a working automatic medication injection mechanism.

Work:

• Design motor-driven multi-dose injection mechanism
• Engineer sterile needle containment for 30-day body wear
• Design drug cartridge with temperature and tamper monitoring
• Test injection reliability across skin types, body compositions, and device orientations
• Validate drug stability in body-worn conditions over 90-day simulated wear

Gate: 99.9% mechanical injection reliability (1,000 bench tests). Drug potency >90% of labeled concentration after 30 days at 37 degrees C.

Cost estimate: $4-8M
Key risk: sterile barrier engineering. Maintaining sterility over 30 days of body wear with sweat, motion, and skin flora exposure.

Phase 4: Integration Prototype (Months 30-48)

Objective: combine all four layers into a single wearable device.

Work:

• Integrate sensor array, decision engine, AMI, and communication into one housing
• Optimize power management and battery life
• Implement security architecture (HSM, encrypted communication, tamper detection)
• Build control hub software and staffing workflow
• Conduct end-to-end system testing (sensor to injection to hub notification)

Gate: device operates for 7+ days on single charge. End-to-end latency from detection to injection <500ms for cardiac events, <60 seconds for respiratory depression. Weight under 120g.

Cost estimate: $5-10M
Key risk: miniaturization. Fitting all components within acceptable body-worn dimensions while maintaining thermal management, signal integrity, and mechanical reliability.

Phase 5: Pre-Clinical Testing (Months 42-60)

Objective: demonstrate safety and efficacy in controlled non-human and simulated-human testing.

Work:

• Animal studies for injection mechanism safety and drug absorption
• Simulated-use testing with human factors study participants (device worn, mock emergencies triggered, no actual injection)
• Biocompatibility testing per ISO 10993
• Electrical safety testing per IEC 60601
• EMC testing for wireless communication interference
• Cybersecurity penetration testing by independent lab

Gate: all ISO/IEC testing passed. Human factors study shows >95% correct device operation by untrained users. No unanticipated adverse effects in animal studies.

Cost estimate: $8-15M
Key risk: FDA may require novel testing protocols for autonomous injection devices. No direct regulatory precedent exists.

Phase 6: Clinical Trials (Months 54-84)

Objective: demonstrate safety and efficacy in human subjects sufficient for FDA approval.

Work:

• Phase I: small cohort (30-50) safety study. Device worn, algorithm runs, injection mechanism tested with saline. Primary endpoint: no device-related adverse events.
• Phase II: larger cohort (200-500) in at-risk populations. Supervised use environment. Real medication loaded. Primary endpoint: detection accuracy in real-world conditions.
• Phase III: pivotal trial (1,000+). Unsupervised home use. Primary endpoint: reduction in adverse outcomes compared to standard of care.

Gate: statistically significant improvement in outcomes with acceptable safety profile. FDA pre-market approval submission.

Cost estimate: $20-50M (depending on trial design and FDA requirements)
Key risk: recruitment. Finding 1,000+ patients at high risk of anaphylaxis, overdose, or MI who will consent to wear an experimental autonomous injection device. Trial design for conditions with unpredictable event timing.

Total Estimated R&D Investment: $42-92M over 7 years

This does not include manufacturing scale-up, marketing, distribution, or ongoing regulatory compliance costs. Total cost to first commercial unit: likely $80-150M.

Closing Note

SmartShot's naloxone variant is the most buildable. The detection problem (respiratory depression) has a proven sensing approach. The medication has high false-positive tolerance. The injection pharmacology is established. A focused program could reach clinical trials within 4 years.

The epinephrine variant depends on solving anaphylaxis detection — a problem that the best lab at Harvard has been working on for years without reaching clinical readiness.

The cardiac variant has the weakest technical foundation. Single-lead MI detection is unreliable, and there is no established subcutaneous medication for acute cardiac events.

The honest assessment: SmartShot is three products with radically different feasibility profiles marketed under one brand. The naloxone variant could exist. The epinephrine variant might exist if biomarker sensing advances. The cardiac variant requires breakthroughs in both sensing and pharmacology that may not arrive this decade.

The control hub is the element that makes any of this defensible. Machines will not be trusted to inject drugs autonomously — not by the FDA, not by patients, not by the liability insurers. Humans in the loop turn an impossible regulatory problem into a difficult one.