Optimizing Machine Learning for High Sensitivity Screening

44 concepts43 papers5 proposals

Standard machine learning models are typically evaluated on symmetric metrics like accuracy or optimized using balanced loss functions like standard cross-entropy. These frameworks implicitly assume that all errors are created equal: misclassifying a positive instance as negative (a False Negative) carries the exact same penalty as misclassifying a negative instance as positive (a False Positive).

In many high-stakes domains, this assumption completely breaks down. Consider clinical diagnostics, software security, or disaster response. If an algorithm flags a healthy patient as having a disease, the cost is a minor inconvenience—a secondary test or human review. But if the algorithm clears a patient who actually has a malignant tumor, the cost is catastrophic. As Amit Seal Ami et al. demonstrate in "False negative - that one is going to kill you": Understanding Industry Perspectives of Static Analysis based Security Testing, software practitioners are often willing to tolerate an 80% false positive rate if it guarantees that a single "killer" vulnerability is not missed.

To bridge this gap, a growing body of research focuses on sensitivity-prioritized machine learning, designing systems that explicitly account for asymmetric error costs to maximize recall.

Mechanisms for Shifting the Objective

To understand why sensitivity-prioritized models work, we must look at how researchers alter the model's optimization landscape to favor recall over precision.

Data-Level and Active Sampling

One intuitive approach is to dynamically alter what the model sees during training. In systematic literature reviews, missing a single relevant study can invalidate an entire review's conclusions. To combat this, Thomas A Trikalinos et al. combine a "Patient Active Learning" strategy with aggressive undersampling of the majority (negative) class in Semi-automated screening of biomedical citations for systematic reviews. This forces the model to continuously refine its decision boundary around rare positive instances, reducing manual screening workloads without missing critical evidence.

Modifying the Latent Space

Instead of relying on standard classification boundaries, models can learn representations that explicitly isolate rare patterns. Krati Gupta et al. employ deep metric learning via a Siamese network with triplet loss to identify rare mitotic patterns in cell images in Deep Metric Learning for Identification of Mitotic Patterns of HEp-2 Cell Images. By optimizing for relative distances in an embedding space rather than a hard binary threshold, the network successfully identifies "needles in a haystack" without requiring artificial data balancing. Similarly, Tianling Liu et al. use task-aware contrastive learning in Joint Prediction of Meningioma Grade and Brain Invasion via Task-Aware Contrastive Learning to disentangle multi-modal MRI features, closing the dangerous sensitivity gap that often leaves aggressive pathologies undetected.

Architectural Inductive Biases

Sometimes the key to high sensitivity lies in forcing the model to look at the right features. In Computed tomography–based COVID–19 triage through a deep neural network using mask–weighted global average pooling, Hongtao Zhang et al. address COVID-19 triage by introducing a mask-weighted Global Average Pooling (GAP) layer. By using a segmentation mask to focus the network's attention strictly on pulmonary tissue, they eliminate background noise that might otherwise cause a model to overlook subtle, early-stage infections.

The Automated Triage and Safety Net Paradigm

Why do these highly sensitive models work so well in practice? They are rarely deployed as standalone decision-makers; instead, they function as an automated triage filter or a safety net.

[ Raw Input Data ] ──> [ High-Sensitivity AI Filter ] ──> [ Enriched Positive Pool ] ──> [ Human Expert Review ]
                                 │                                     
                                 └──> [ Confirmed Negatives ] (Safely Dismissed)

In this paradigm, the machine learning model acts as the first line of defense. Its job is not to be perfectly precise, but to be perfectly comprehensive. It filters out the obvious negatives, leaving a significantly smaller, enriched pool of potential positives for human experts to review.

For example, Christina Fell et al. developed a two-stage pipeline to triage endometrial cancer biopsies in Detection of malignancy in whole slide images of endometrial cancer biopsies using artificial intelligence, achieving 97% sensitivity to ensure high-risk patients are prioritized immediately. In a massive real-world deployment screening over 600,000 patients for diabetic retinopathy, Arthur Brant et al. demonstrated a 0% "clinically important miss rate" in Performance of a Deep Learning Diabetic Retinopathy Algorithm in India, proving that automated triage can safely eliminate human workload without sacrificing patient safety.

This safety net effect is also highly effective in time-critical or resource-constrained settings:

Cross-Domain Imperatives of High Recall

The mathematical and structural need for sensitivity-prioritized learning extends far beyond oncology and clinical imaging.

Crisis Intervention

In suicide prevention, a false negative is fatal. Meghan Broadbent deployed a transformer-based RoBERTa model to analyze text-based crisis counseling encounters in A machine learning approach to identifying suicide risk among text-based crisis counseling encounters. By prioritizing sensitivity, the model catches subtle linguistic risk signals that human counselors might overlook under high-pressure conditions.

Disaster Response

When deploying models to edge hardware, such as low-cost orbital satellites, Gonzalo Mateo-Garcia et al. optimize lightweight CNNs for flood detection in Flood Detection On Low Cost Orbital Hardware. Because failing to detect a disaster has severe humanitarian consequences, the system is tuned for maximum coverage and sensitivity, accepting higher false-positive rates as a necessary trade-off for real-time orbital triage.

Data Privacy

When de-identifying patient notes to protect privacy, Ji Young Lee et al. demonstrate in Feature-Augmented Neural Networks for Patient Note De-identification that pure neural networks must be augmented with structured metadata to maximize recall. Failing to catch a single sensitive identifier constitutes a legal and ethical breach, making a high-recall focus mandatory.

The Inherent Tension of the False Positive Burden

While prioritizing sensitivity protects against catastrophic failures, it introduces a fundamental operational challenge: the false positive burden.

As a model's decision threshold is lowered to capture every possible true positive, the number of false alarms naturally rises. As Alison O’Mara-Eves et al. point out in Erratum to: Using text mining for study identification in systematic reviews: a systematic review of current approaches, achieving near-perfect recall inevitably forces humans to manually screen a massive volume of false positives. The technical challenge is not just maximizing sensitivity, but doing so while minimizing this manual screening burden.

Conversely, there are high-stakes scenarios where the asymmetry is inverted, revealing the delicate nature of threshold tuning. Felix Krones et al. study the prediction of neurological recovery from comas after cardiac arrest in Multimodal Deep Learning Approach to Predicting Neurological Recovery From Coma After Cardiac Arrest. In this environment, a false positive—predicting a patient will not recover when they actually could—can lead to the premature withdrawal of life support. Because the cost of a false positive is an irreplaceable human life, the decision thresholds must be kept exceptionally strict, which inherently restricts the model's sensitivity and makes it harder to identify the broad range of patients who might ultimately recover.

Key Foundational Work

Semi-automated screening of biomedical citations for systematic reviews (2010) — contributed 1 concept elements to this direction.
Boosting Face Identification in Airports (2003) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_asymmetric_error_costs_27c8449b,
  title = {asymmetric error costs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-error-costs/rdr_27c8449b2ae34d68a69afe117d8c1127},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_asymmetric_error_costs_27c8449b_v1,
  title = {asymmetric error costs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-error-costs/rdr_27c8449b2ae34d68a69afe117d8c1127/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake