Lasers tuned to the strongest infrared absorption line of water, the Er:YAG laser at 2,940 nm, remove tissue by flash-boiling the water trapped inside it. The trapped water turns to steam, the tissue breaks apart in tiny explosions, and most of the heat leaves with the ejected debris. Dentists use these lasers to strip bacterial deposits off tooth roots, and the same micro-explosions kill bacteria at energies far below those needed to cut hard deposits. The lasers have one large clinical weakness: they never touch the tooth, so the operator cannot feel when the diseased layer is gone and healthy root surface has been reached. Each micro-explosion does make a sound, though. Recent work showed that those sounds carry information about what material was just hit. We propose to train a model that listens to the pulse-by-pulse sound and decides when to stop. The new element is how the "right answer" is defined. A spot counts as done when sequencing of the bacterial 16S rRNA gene says no living bacteria are left, where 16S is the ribosomal gene used as a universal DNA barcode for identifying bacteria. Visual cleanliness and tissue type do not set the endpoint.
The sequencing does more than supply labels. It tests a hypothesis that, if true, makes a simple biomass detector unsafe. Bacterial species may differ in how readily laser micro-explosions kill them, so the community left behind after partial treatment would not be a scaled-down copy of the original. It could be enriched for the harder-to-kill and possibly more harmful members. If so, a stop rule tuned only to total bacterial load would stop too early in exactly the cases that matter. The sequencing must also count only living cells. Laser-lysed bacteria leave their DNA behind, and ordinary 16S sequencing would read that DNA as survivors.
Core Claim
The acoustic emission of an Er:YAG pulse train contains enough information to predict whether a treated spot still carries viable bacteria, and in particular viable members of taxa that survive laser exposure. A stop rule built on that prediction will leave fewer live pathogens and remove less healthy root than a fixed-dose clinical protocol. Two things make this testable and new. Ground truth is defined by viability-filtered, full-length 16S profiling rather than by tissue class. The analysis also asks whether species differ in their laser dose–response, which determines whether composition-aware labels are needed at all.
What Prior Work Establishes and Leaves Open
Three strands motivate the proposal.
The first is about killing bacteria. Bactericidal Effect of Er:YAG Laser on Periodontopathic Bacteria. irradiated pure cultures of P. gingivalis and A. actinomycetemcomitans on agar and found growth inhibition from 0.3 J/cm², well below the 7.1–10.6 J/cm² used for clinical calculus scaling. It counted colonies of two culturable species, so it says nothing about mixed biofilms or about whether some species tolerate the laser better than others. We extend that dose–response to whole communities.
The second is clinical. Reviews of periodontal Er:YAG use, Dental Calculus Removal with Lasers and Current Status of Laser Application in Periodontics, document three points:
- hard deposits ablate selectively over healthy cementum because they hold more water;
- thermal damage stays within a shallow zone;
- loss of tactile feedback is the main practical barrier.
Their microbiology relies on cultures or probes for a few named pathogens. Our endpoint is defined over the whole community.
The third strand concerns what sequencing can see. Microbial predictors of healing and short-term effect of debridement on the microbiome of chronic wounds sequenced the V1–V3 region of 16S in chronic wounds before and after scalpel debridement. It found a weighted UniFrac distance of only 0.086, where weighted UniFrac is a community dissimilarity score that accounts for how related the differing taxa are. It also linked facultative anaerobes such as Enterobacter to non-healing. We read this paper as a warning as well as motivation. Its sequencing did not separate living from dead cells, and any study of a killing intervention must. Evaluation of 16S rRNA gene sequencing for species and strain-level microbiome analysis supplies the sequencing method. It showed that the common V4 sub-region leaves 56% of sequences unclassified at species level. Full-length PacBio reads, together with the small differences between a genome's multiple 16S copies, resolve species and even strains. We need that resolution because species-specific survival is the hypothesis under test.
For sensing, Deep-Learning Approach for Tissue Classification using Acoustic Waves during Ablation with an Er:YAG Laser (Updated) supplies both the method and the ceiling. That work recorded 25,000 pulses on five sterile porcine tissues with an air-coupled piezoelectric microphone. A hybrid model combining a convolutional network on the waveform with a fully connected network on its spectrum reached 65.5%–75.5% accuracy, and its attribution maps pointed to low frequencies. It never tested biofilm or calculus, and it classifies single pulses with no memory of earlier pulses at the same spot. Deep learning models comparison for tissue classification using optical coherence tomography images: toward smart laser osteotomy supplies the strongest competing modality. Optical coherence tomography (OCT) is depth imaging with reflected infrared light. A DenseNet161 model on OCT patches reached 94.85% accuracy and 94.67% F1 on the same kind of porcine tissue task. Finally, Learning, visualizing and exploring 16S rRNA structure using an attention-based deep neural network shows that neural models can read raw 16S sequences directly. We do not need that capability for labeling and use standard reference-based classification instead.
Research Program
Specimens. The bench system uses two kinds of specimen.
- Extracted human teeth carrying natural subgingival calculus, with each donor tooth kept as a unit for cross-validation.
- Root slabs coated with multispecies biofilm grown from pooled subgingival plaque, plus a defined oral consortium of known composition that includes P. gingivalis and A. actinomycetemcomitans. The defined consortium acts as a mock community in the sense used by Johnson et al.: the true composition is known, so sequencing bias can be measured.
Laser and recording. We irradiate with an Er:YAG handpiece across a fluence grid running from the 0.3 J/cm² bactericidal threshold to the 10.6 J/cm² scaling range. Repetition rate is 20–30 pulses per second, with and without coaxial water spray. Every pulse is recorded with an air-coupled piezoelectric transducer configured as in the 2024 acoustic study, using a 384 µs window.
Survival-curve design. Sampling destroys a spot, so each specimen is divided into many replicate spots, and each spot is stopped after a different number of pulses. The collection of spots then traces a survival curve. After treatment, each spot is microbiopsied and split three ways:
- total-DNA full-length 16S,
- full-length 16S after propidium monoazide treatment, a dye that blocks amplification of DNA from membrane-compromised cells, so that only intact cells are counted,
- quantitative PCR (qPCR) of total 16S copies, which measures absolute load.
A cross-section of each spot is imaged to measure how much healthy cementum was removed.
Analyses. The first analysis is ecological. For each species, we fit live-cell abundance against cumulative fluence and ask whether the decay rates differ beyond what sequencing noise allows. The defined consortium calibrates that noise. We model donor and specimen as random effects, following the hierarchical approach Verbanic et al. used for patient heterogeneity. The second analysis is the detector. A sequence model reads the acoustic history of a spot, meaning every pulse so far and not just the latest. At each pulse it predicts two quantities: residual viable load, and a pathogen-weighted residual score. That score weights live abundance of taxa found in the first analysis to survive laser exposure and of named periodontal pathogens. The stop rule fires when the predicted residual falls below a set threshold.
Baselines
The naive reference is the fixed-dose protocol clinicians use today. The operator applies a set number of passes at scaling fluence with water spray. We run it on matched spots from the same donor teeth and score it on residual live pathogen load and healthy-root loss. Beating it means lower pathogen-weighted residual at equal or lower cementum loss. We also report total pulses delivered, since fewer pulses means shorter procedures.
The strongest existing acoustic method is the hybrid CNN+FcNN of the 2024 acoustic study, which reported 65.5%–75.5% accuracy on five sterile porcine tissue classes. We retrain it on our recordings with the same donor-held-out splits and the same hyperparameter search budget as our model. Its task is reframed as single-pulse classification of "deposit or biofilm present" versus "clean substrate", with the target defined by our viability labels. There is no reported result found on Lacuna for acoustic detection of biofilm or calculus, so this retrained model is the comparison number. Our sequence model beats it if it improves AUROC (area under the curve plotting true- against false-positive rate) on the pathogen-weighted endpoint on held-out donors. It also needs to reduce stop-point error, measured as the number of pulses between predicted and true clearance.
The strongest competing modality is OCT with DenseNet161, the model that reached 94.85% accuracy and 94.67% F1 on tissue patches in the OCT osteotomy study. We image the open-surface specimens before and after each pulse block, train on the same labels and splits, and compare the two modalities on the same spots.
Three ablations isolate the ingredients:
- the same sequence model trained on total-DNA labels rather than live-cell labels, which tests whether viability filtering matters;
- the same model trained on qPCR load alone, which tests whether composition matters;
- a model with no pulse history, which tests whether trajectory matters.
Why not just…?
Why not just use OCT, which already reaches about 95%? On open surfaces OCT may well win, and we will report that honestly. Two arguments limit its reach. OCT sees structure, and a thin, structurally modest biofilm or a viable layer on the surface of calculus need not have an optical boundary. OCT also needs a clear line of sight, which deep periodontal pockets, blood, and the debris plume deny it, the access problem stressed in Current Status of Laser Application in Periodontics. The head-to-head on open specimens settles the first argument. A second comparison under saline and blood-tinged fluid films probes the second.
Why not just label with qPCR of total bacterial load, or with culture counts? This is the real threat. If every species dies at the same rate under the laser, total live load is a sufficient endpoint, and full-length 16S adds cost without adding information. Culture is weaker still, because it counts only culturable species and only the few that were plated for, as in Ando et al.. The first analysis decides the question directly. If species-specific decay rates are indistinguishable, we drop composition from the endpoint and keep qPCR plus viability filtering. The qPCR-label ablation then shows whether the detector loses anything.
Why not just apply a generous fixed dose, given that bacteria die at 0.3 J/cm²? The bactericidal threshold was measured on agar monocultures. Bacteria embedded in calculus and matrix are shielded by the deposit itself, so the dose actually reaching them is unknown. Over-treating spends healthy cementum, which is the very tissue the selective-ablation argument in Dental Calculus Removal with Lasers is meant to protect. The fixed-protocol baseline, run at several pass counts, maps the trade-off between residual pathogens and cementum loss. The detector has to lie below that curve.
Proposed Experiments and Decisive Checks
The first check is the viability control. On the defined consortium, we expect live-cell sequencing to show steep loss of viable bacteria with fluence while total-DNA sequencing changes much less. That result would demonstrate that ordinary 16S profiling misreads killing interventions. It would also motivate repeating debridement comparisons such as Verbanic et al.'s with viability filtering. If the two readouts agree, the viability arm is unnecessary.
The second check is differential susceptibility. Using full-length reads resolved to species, we test whether laser decay rates differ between taxa, including whether facultative anaerobes and named pathogens decay more slowly. A null result simplifies the endpoint, as described above.
The third check is the detector. On donor-held-out teeth, the acoustic sequence model must beat the fixed protocol on residual pathogens at matched cementum loss. It must beat the retrained single-pulse CNN+FcNN on AUROC and stop-point error. It must also show that pulse history carries the signal. If acoustic AUROC stays near the porcine-tissue ceiling while OCT clears the task on open surfaces, the conclusion is that acoustics is useful only where optics cannot reach. We would then focus the work on the fluid-film comparison. All acoustic recordings, paired with their sequencing labels, will be released as the first sound-to-microbiology dataset for laser debridement.
References
- Bactericidal Effect of Er:YAG Laser on Periodontopathic Bacteria.
- Dental Calculus Removal with Lasers
- Current Status of Laser Application in Periodontics
- Microbial predictors of healing and short-term effect of debridement on the microbiome of chronic wounds
- Evaluation of 16S rRNA gene sequencing for species and strain-level microbiome analysis
- Deep-Learning Approach for Tissue Classification using Acoustic Waves during Ablation with an Er:YAG Laser (Updated)
- Deep learning models comparison for tissue classification using optical coherence tomography images: toward smart laser osteotomy
- Learning, visualizing and exploring 16S rRNA structure using an attention-based deep neural network