PCR Signal Baseline Correction

Redesigning a hard-coded legacy baseline algorithm into a data-driven mixed-basis-function model that cut the false-negative rate by 91%.

English한국어

Role: Technical Lead / Data Scientist (Data Scientist 3, Data Engineer 1)  ·  Period: 2024.01 – 2024.09  ·  Stack: Python, Matlab, mixed-basis-function modeling, lightweight linear regression

Led the redesign of a hard-coded legacy PCR signal-correction algorithm into a data-driven mixed-basis-function model. The legacy stack was brittle for two reasons:

  • Signal complexity — chemical, optical, and mechanical responses mix, so no single deterministic rule covers it; conditional branches piled up for every new noise pattern.
  • Fragmented tooling — multiple non-standard baseline-fitting algorithms coexisted, giving inconsistent results across products and teams.
PCR Signal Baseline Correction — project poster

Highlights

  • Data-driven mixed-basis-function model — inspired by Taylor-series polynomial approximation, I designed a “characteristic-equation” algorithm that transforms the signal through mixed polynomial / exponential / log basis functions and fits the baseline with lightweight linear regression. Adding a basis function extends it to a new signal pattern — no new conditional branches.
  • False-negative rate 0.47% → 0.04% (91.49% improvement) — cutting patient-safety-critical false negatives to a fraction of their prior level.
  • Ranked #1 in residual-signal white-noise approximation against 5 competing algorithms, including a head-to-head comparison with the industry-leading third-party black-box algorithm.
  • Minimal-package constraint — implemented without external ML/DL frameworks (NumPy / Pandas only) to keep the pipeline portable to C++, with an optimization pipeline of log normalization → basis-function feature vectors → cost function → momentum gradient descent → prediction → denormalization.
  • Intuitive comparison dashboards (multi-signal, single-signal, per-signal-type) let biologists and executives take part in objective, evidence-based decisions.

Approach

A single measured, data-driven model replaces a hard-coded baseline, unifying fragmented fitting algorithms and validating quality against competing methods on the same signals.

flowchart TB
    L[Hard-coded legacy algorithm] --> D[Data-driven<br/>mixed-basis-function model]
    D --> OPT[Log norm, basis vectors,<br/>momentum gradient descent]
    OPT --> CMP[6-algorithm comparison<br/>+ visual validation]
    CMP --> FN[False-negative<br/>0.47% to 0.04%]
    CMP --> WN[#1 white-noise<br/>approximation of 5]