Falsification & Open Questions

This document is not pessimism. It is due diligence written before anyone asks for it. Every question here is answerable. Answering them is the work. But pretending these questions don't exist would be malpractice.

1. Technical Risks That Could Kill the Concept

Sensor accuracy is the first wall. Current wearable biosensors can detect opioid-induced respiratory depression with roughly 90% accuracy using photoplethysmography. That number comes from supervised injection facility studies — controlled environments with clinical oversight. Ninety percent is remarkable for research. It is nowhere near sufficient for autonomous injection.

The threshold question: what false-positive rate is acceptable for a device that injects medication without human confirmation? One false injection per thousand hours of wear? Per ten thousand? Per hundred thousand? I don't know the answer yet. Nobody does. The answer determines whether the concept is viable at the sensor level.

Anaphylaxis detection faces a different problem. Epinephrine auto-injection requires distinguishing anaphylaxis from a panic attack. Both present with tachycardia, diaphoresis, and subjective distress. The physiological signatures overlap in ways that current wearable sensors cannot reliably separate. A false epinephrine injection during a panic attack is not life-threatening, but it is medically harmful and legally catastrophic.

Drug stability is the second wall. Epinephrine degrades with exposure to heat, light, and oxygen. Current storage recommendations require controlled room temperature: 20-25 degrees Celsius, with excursions permitted to 15-30 degrees. A body-worn device operates at skin temperature — approximately 33-36 degrees Celsius — and subjects the drug reservoir to continuous movement, vibration, and sweat exposure.

EpiPens already struggle with this. Autoinjectors stored at 70 degrees Celsius for 10 days retain only 63% of labeled epinephrine concentration. At body temperature, degradation is slower but persistent. Autoinjectors up to 30 months past expiration still show 90% concentration, which suggests the timeline is workable — but in a sealed, static device. A body-worn reservoir with repeated thermal cycling and mechanical stress is a different engineering problem.

Naloxone is more thermally stable than epinephrine. Nitroglycerin is less stable. Each drug in the SmartShot portfolio has a different stability profile, and each requires its own engineering solution.

Miniaturization is the third wall. The full system requires biosensors, onboard compute, a drug reservoir, an injection mechanism, a battery, and cellular communication — all in a form factor that a person will actually wear. The closest analog is the insulin pump. A Medtronic MiniMed 780G weighs approximately 100 grams and delivers a single drug through a single mechanism. SmartShot requires more sensing, more compute, and a fundamentally different injection system.

Battery life compounds the miniaturization problem. Continuous biosensing plus cellular communication plus injection mechanism actuation creates a power budget that current lithium-polymer cells struggle to meet in a wearable form factor. If the device requires daily charging, compliance drops. If compliance drops, the device fails its core promise. This is a circular engineering constraint.

2. Regulatory Risks

The FDA has never cleared a device that autonomously delivers emergency medication based on algorithmic detection of a medical event. The closest precedent is the Medtronic MiniMed 670G, cleared in 2016 as the first hybrid closed-loop insulin delivery system. But insulin delivery in a diabetic is a maintenance function. Emergency medication delivery in a crisis is categorically different.

Each drug-condition combination will likely require a separate Premarket Approval. Four conditions — opioid overdose, anaphylaxis, cardiac emergency, diabetic crisis — could mean four separate PMAs. Average cost to bring a device through PMA is approximately $94 million. Four PMAs could exceed $400 million in regulatory costs alone, before manufacturing, marketing, or distribution.

The FDA's regulatory framework for AI/ML-based medical devices is evolving rapidly. As of early 2026, the FDA has authorized over 1,350 AI-enabled devices, but the vast majority are diagnostic — they inform human decisions rather than execute autonomous treatment. The agency published draft lifecycle management guidance in January 2025 and finalized predetermined change control plan guidance in August 2025. These frameworks address algorithms that change over time, but they do not address algorithms that autonomously deliver medication.

The regulatory risk is not that FDA will reject SmartShot. The risk is that FDA will determine the device requires a regulatory framework that does not yet exist. Creating new frameworks takes years. Sometimes decades. A device that sits in regulatory limbo is a device that burns capital without generating revenue.

Post-market surveillance requirements add another layer. A device that autonomously injects medication will require continuous real-world performance monitoring at a scale and granularity that few medical devices face. The cost of that surveillance infrastructure is unknown and potentially prohibitive.

3. Liability — The Unanswered Question

This is the risk I am least able to quantify.

If SmartShot injects epinephrine during a panic attack — not anaphylaxis — who is liable? The manufacturer, for the algorithm's misclassification? The prescribing physician, for authorizing autonomous delivery? The Control Hub operator, for not overriding? No existing liability framework provides a clear answer.

If SmartShot fails to inject during a real emergency and the patient dies, that is a product liability case with direct precedent. The EpiPen litigation is instructive. By 2017, the FDA had received 720 reports of EpiPen malfunctions, including over 200 hospitalizations and at least seven deaths. The manufacturer faced class action lawsuits for devices that failed to deploy needles or emptied medication before injection. SmartShot would face the same exposure, amplified by the fact that the device promises autonomous protection.

If a patient removes or disables the device and then has an emergency, the liability question inverts. Does the manufacturer bear any responsibility for not preventing removal? For not alerting a caregiver? The answer depends on the use case. A harm reduction patient who removes the device voluntarily is different from a court-mandated patient who removes it to use opioids.

The Medtronic insulin pump litigation offers a parallel. The MiniMed 600 series generated over 26,000 error reports, more than 2,000 injuries, and at least one death. The Class I recall found that pumps were incorrectly dosing insulin. One plaintiff alleged a 630G pump delivered its full insulin payload at once, causing a diabetic coma. These cases establish that autonomous medication delivery creates manufacturer liability when the autonomy fails.

But SmartShot's liability exposure is broader than Medtronic's. Insulin pumps deliver a drug the patient already takes daily. SmartShot delivers emergency drugs in crisis situations to people who may be unconscious. The legal exposure is genuinely uncharted.

No existing liability framework covers autonomous emergency medication delivery. Medical malpractice law assumes a human decision-maker. Product liability law assumes a product that performs a defined function. SmartShot sits between these — a product that makes clinical decisions. Recent legal scholarship has begun addressing this gap, but the law has not caught up. Courts and legislatures will define the framework reactively, through litigation. That means the first manufacturer to market absorbs the full cost of establishing legal precedent.

4. Ethical Questions

The correctional and court-mandated use case is the one that keeps me up at night.

Is auto-injection of psychiatric medication in a correctional setting treatment or coercion? The Constitution permits involuntary administration of antipsychotic drugs only under narrow conditions established in Washington v. Harper: the treatment must be medically appropriate, substantially unlikely to have side effects that undermine fairness, and significantly necessary to further important governmental interests. A wearable device that continuously monitors and autonomously injects raises questions that Harper did not contemplate.

Consent withdrawal is a related problem. Can a patient disable the device during a detected emergency? Should they be able to? If the device detects an opioid overdose and the patient — still conscious — removes it before injection, is that an exercise of autonomy or a symptom of the emergency? I do not have a clean answer.

Data ownership is a quieter but equally consequential question. Continuous vital sign monitoring generates a detailed physiological record. Who owns it? The patient? The device manufacturer? The insurer who subsidizes the device? If a harm reduction program provides SmartShot and the device records every use event, does that data become evidence in a criminal proceeding?

Algorithmic bias is a risk in any ML-based diagnostic system. If detection algorithms are trained on datasets that underrepresent women, Black patients, or patients with high BMI, the device will fail disproportionately in those populations. Given that the opioid crisis disproportionately affects specific communities, algorithmic bias in SmartShot would compound existing health disparities.

5. Competitive Threats

Apple has wrist-worn biosensors, health AI infrastructure, and distribution to hundreds of millions of users. If Apple decides to build drug delivery into Apple Watch, SmartShot becomes irrelevant overnight. Apple has not signaled interest in drug delivery, but they have not signaled disinterest either. Their biosensor capabilities grow with every generation.

Abbott has continuous biosensing infrastructure through FreeStyle Libre. Medtronic has closed-loop drug delivery infrastructure through its insulin pump portfolio. Either company could pivot to emergency medication delivery using existing manufacturing, regulatory relationships, and clinical evidence. They have not done so. That could mean the market is unattractive, or it could mean the timing is not right.

The University of Washington published proof-of-concept results for a wearable naloxone injector in Scientific Reports in November 2021. The device uses accelerometers to measure respiration, detects apnea, and injects naloxone subcutaneously. It works. It was tested in a supervised injection facility in Vancouver and a clinical trial in Seattle. If that team secures funding and advances to formal clinical trials, they will have a years-long head start in the specific use case — opioid overdose — where SmartShot has the clearest market.

A well-funded startup with backing from a16z bio, ARCH Venture, or a similar life sciences investor could move faster with more capital. SmartShot's advantage is the thesis, not the technology. Theses can be replicated.

6. The Seven Things We Don't Know

One. Whether continuous wearable sensors can achieve the accuracy required for autonomous injection safety — not 90%, but 99.9% or better — across diverse populations and real-world conditions.

Two. Whether target medications remain stable in a body-worn form factor through thermal cycling, mechanical stress, and sweat exposure over clinically meaningful timeframes.

Three. Whether the FDA will clear an autonomous emergency medication delivery device under any existing pathway, or whether a new framework is required.

Four. Whether the liability question is solvable before the first wrongful-injection or failure-to-inject lawsuit defines the legal landscape through litigation.

Five. How much it would actually cost to bring SmartShot from concept to first FDA clearance. The honest range is $50 million to $200 million, and that spread tells you how much uncertainty remains.

Six. Whether the first customer is a harm reduction program, a health system, a correctional facility, or a military contract — and how that choice shapes the product, the regulatory strategy, and the ethical posture of the company.

Seven. Whether any of this matters if a major player builds it first. SmartShot is a thesis. A thesis without execution is a Wikipedia article.

Closing Note

Every question in this document is answerable. Some require engineering. Some require clinical trials. Some require regulatory negotiation. Some require litigation. None of them are reasons to stop. All of them are reasons to be honest about what we know and what we don't.

The structural honesty in this document is what separates a thesis from a fantasy. I would rather list every way this can fail than pretend the path is clear. The path is not clear. The destination is worth the difficulty of finding it.