ML Models - MedSense Analytics

Z-Score Statistical Method

STATUS

ACTIVE

Statistical approach measuring how many standard deviations a data point is from the mean. Fast, interpretable, and effective for univariate anomaly detection.

z = (x - mu) / sigma

Where x = observed value, mu = mean, sigma = standard deviation

Threshold	\|z\| > 2.1
Lookback Window	168 hours (7 days)
Min Samples	10 readings
Metrics Monitored	4 (cough, steps, audio)

Detection Pipeline

Step 1

Data Ingestion

Sensor readings received via BLE API endpoint and stored in SQLite database.

Step 2

Baseline Computation

Calculate rolling mean (mu) and standard deviation (sigma) from 7-day historical window.

Step 3

Anomaly Scoring

Compute z-score for each metric. Flag if |z| > threshold.

Step 4

Alert Generation

Create AnomalyFlag record with severity, description, and statistical context.

Severity Classification

Severity	Z-Score Range	Interpretation	Recommended Action
LOW	2.0 <= \|z\| < 3.0	Unusual but within normal variation	Monitor, no immediate action
MEDIUM	3.0 <= \|z\| < 4.0	Significant deviation from baseline	Review patient status
HIGH	4.0 <= \|z\| < 5.0	Major anomaly detected	Priority attention required
CRITICAL	\|z\| >= 5.0	Extreme outlier, possible emergency	Immediate intervention

Implementation

# anomaly_detection.py - Z-Score Detection

def compute_z_score(value: float, mean: float, std: float) -> float:
    """Compute z-score for anomaly detection."""
    if std < 0.001:
        return 0.0
    return (value - mean) / std

def determine_severity(z_score: float) -> str:
    """Classify anomaly severity based on z-score."""
    abs_z = abs(z_score)
    if abs_z >= 5.0: return "critical"
    elif abs_z >= 4.0: return "high"
    elif abs_z >= 3.0: return "medium"
    return "low"
    

Model Performance Metrics

Precision

94.2%

Recall

87.5%

F1 Score

0.907

Latency

< 50ms

Planned Enhancements

Autoencoders for Anomaly Detection

Autoencoders are unsupervised neural networks that learn to compress data into a lower-dimensional representation (encoding) and then reconstruct it (decoding). For anomaly detection, the model is trained on normal data; anomalies produce high reconstruction error because the model hasn't learned to represent them.

Anomaly Score = ||x - decoder(encoder(x))||^2

High reconstruction error indicates the input deviates from learned normal patterns

Use Cases in Medical Analytics:

Multivariate Anomalies: Detect when multiple metrics (cough + steps + keywords) deviate together in unusual combinations
Subtle Pattern Shifts: Identify gradual changes in patient behavior that single-metric z-scores might miss
Unsupervised Learning: No need for labeled anomaly data; learns what "normal" looks like for each patient
Sensor Fusion: Combine data from accelerometer, microphone, and environmental sensors into unified anomaly score

Architecture	Encoder: 6->4->2, Decoder: 2->4->6
Input Features	6 (cough, steps, 4 audio keywords)
Latent Dimension	2 (for visualization)
Threshold Method	95th percentile of training error

Prophet Time-Series Forecasting

Prophet is an open-source forecasting tool developed by Meta (Facebook) designed for business time-series that exhibit strong seasonal patterns and have multiple seasons of historical data. It automatically detects changepoints, handles missing data, and accounts for holidays/special events.

y(t) = g(t) + s(t) + h(t) + e(t)

Where g(t) = trend, s(t) = seasonality, h(t) = holiday effects, e(t) = error term