Ethics · Fairness · Limits
Sophisticated scientists and physicians are wary of promises of "technological perfection" in mental health. Demonstrating that FROID's engineering anticipates, quantifies, and mitigates its own biases is our greatest credibility differentiator — and our legal shield.
Pillar 1 · Non-negotiable guidelines
FROID establishes at its foundation that technology expands the therapist's perception without ever assuming the clinical decision.
The system is categorically a coprocessor of biological signals, never an autonomous diagnostic instrument. Assessment and decision-making remain under the ethical and clinical responsibility of the licensed mental health professional.
The patient has the right to know that an AI system is being used in a complementary way in their care, how the data is processed, and what the software's limitations are.
The evolution of the platform's algorithms and practices is structured for periodic review by an independent board including psychologists, psychiatrists, and AI ethics specialists. The board's actual composition is published on About & Contact once constituted.
Pillar 2 · Algorithmic fairness
Every computer vision model or vocal biomarker carries known risks of demographic bias (skin tone, gender, age, accent) and hardware influence. FROID's main technical response is the 60-second Dynamic Baseline.
Neutralizing biological and geographic bias. FROID rejects absolute tables of emotion or vocal tone. Instead of comparing the patient to a generic global standard — which would introduce strong skin-tone bias for computer vision, or accent/language bias for vocal biomarkers — the system establishes a personalized "biological fingerprint" for that patient, on that exact day and time.
Physical-environmental filter. In the first 60 seconds, the bioacoustic engine maps the teleconsultation's background noise and the microphone's frequency response; simultaneously, the visual mapper records the room's lighting and the static muscular characteristics of the resting face (neutralizing age-related sagging or birth asymmetries).
Every analysis measures exclusively the relative mathematical deviation of the patient from their own baseline:
This way, a naturally low voice is not read as motor retardation, and skin under unfavorable lighting does not generate false Action Unit readings — the system operates on the normalized starting state of that same session. Full math in Technology.
Scientific honesty: relative calibration reduces — but does not eliminate — bias. Factors such as extreme signal quality and capture differences remain real limitations, addressed by the Quality Gates (Pillar 3) and continuous auditing (Pillar 5).
Pillar 3 · Product
In teleconsultations, packet loss and hardware variation are critical. Instead of inferring over degraded signal, the system maintains a real-time triage and telemetry layer.
Via MediaPipe, the system detects obstructions and extreme lighting problems (forehead and eyebrows not visible, eyes not visible, etc.). Below a confidence threshold, the dashboard shows a subtle alert: "Low visual confidence — adjust lighting or patient position".
The system monitors the harmonic-to-noise ratio (HNR) and spectral flux. If the connection fluctuates or the microphone distorts in a way that compromises MFCC7/MFCC9, the vocal channel's dynamic reliability factor (rvocal) is reduced in the IPM formula, preventing the indicator from "drifting" due to network noise.
Pillar 4 · Legal compliance
For legal certainty and informational clarity — in compliance with Brazil's LGPD (Law 13.709/2018) and the resolutions of the CFP and CFM (Brazilian psychology and medicine councils) — the notices below are displayed mandatorily in the interface and in reports.
Pillar 5 · Data sovereignty
Edge processing. A meaningful share of phonation and facial-muscle-imaging calculation happens in the browser, on the practice's own computer. In remote sessions, audio and video travel over the internet (with a TURN server relaying traffic when a direct connection isn't possible), and audio segments may be sent to the backend and to a transcription provider for processing — never hidden, always covered by the patient's authorizations. See the full operation in Security.
Granular consent and right to erasure. Audio, video, or transcription capture each require specific, voluntary, and separate authorizations. The patient can revoke consent and request data deletion at any moment; the request is reviewed and may result in deletion, anonymization, blocking, or restricted retention when a legal or professional retention obligation exists (LGPD article 18). We do not promise automatic deletion on a fixed timeline when that obligation exists.
The IPM and IDM weights and thresholds are presented honestly as qualified engineering parameters — proprietary heuristics, not immutable biological truths. Their validation uses the accumulation of anonymized data for demographic deviation studies (for example, auditing whether the ΔΔMFCC9 transition pattern holds across age groups and regional accents within Brazil).
Analytical queries over the Data-Froid require cohorts of at least k = 50 records (k-anonymity), technically enforced on the backend — below that threshold, the query is refused. This substantially reduces, but does not fully eliminate, re-identification risk. See the architecture in Security.
Compliance kit
A base template for adaptation by the professional, aligned with LGPD and CFP/CFM resolutions. Fields in blue must be filled in. This material is a support resource and does not replace guidance from your legal counsel/DPO.
I, [patient's name], authorize [professional's name and license — CRP/CRM] to use, during my care, the FROID system as a tool to support clinical perception.
I declare that I am aware that FROID analyzes voice and facial expression signals to generate statistical indicators supporting the professional, and that it does not perform automated diagnosis nor replace clinical assessment. The therapeutic decision rests entirely with the professional.
I authorize, separately and voluntarily (check):
I have been informed that biometric processing takes place locally on the practice's equipment and that the data is treated as sensitive health data under Brazil's LGPD (Law 13.709/2018).
I am aware that the algorithms have limitations, including lower accuracy for certain demographic groups and sensitivity to lighting and audio conditions, and that results are estimates, not absolute measurements.
I may revoke this consent at any time, without prejudice to the continuity of my treatment, by requesting permanent deletion of the data (LGPD article 18).
I am aware that FROID does not monitor emergencies and that risk situations must be handled through conventional channels (SAMU 192, CVV 188, or local equivalents).
Place and date: [__________] · Patient's signature: [__________] · Professional's signature: [__________]
By adopting this stance, FROID does not promise magic: it delivers high-precision engineering based on relative calibration, with verifiable legal and methodological compliance safeguards. Acknowledging limits with transparency is what positions FROID before enterprise clinics, research ethics committees, and regulatory boards.