Listening to the Doctor: Physician Vocal Adaptation as an Early Signal of Antibiotic Capitulation

Unnecessary antibiotic prescriptions add to resistance, and many of them are written for social reasons rather than medical ones. In a corpus of recorded pediatric consultations, caregivers who pushed for antibiotics raised the odds of a prescription more than ninefold (COSTA). Research on interlocutor adaptation offers a way to see this pressure while it is still building. That field studies how a trained conversational partner quietly changes their own voice, timing and word choice to cope with a difficult counterpart. In autism assessments, the psychologist's speech tracked the child's severity better than the child's own speech did (Bone et al.).

We propose to point the same sensor at the prescribing physician. The claim is that a doctor's vocal adaptation to a caregiver during history-taking predicts whether the visit will end in a guideline-inappropriate antibiotic prescription. Adaptation here means how their pitch, energy, speaking rate and response latency converge toward or pull away from the caregiver's. The prediction should hold before the treatment discussion begins, and it should work better than what the words of the conversation alone reveal. If this is true, a stewardship prompt could reach the physician before they commit to a recommendation. At that point the prompt could still change the outcome, instead of flagging a prescription that is already written.

Core Claim and What Is New

The target is not bacterial resistance. It is the physician's yielding, which is the behavioral step that feeds resistance. We make this choice deliberately. The tempting version of this intersection is to predict carbapenem or piperacillin-tazobactam resistance in bloodstream infections from the clinician's voice. We reject it, because nothing in a doctor's prosody plausibly encodes whether a plasmid carries a carbapenemase gene. The work on bloodstream infections shows that the limit on resistance prediction comes from microbiology and chance, not from missing conversation (Machine learning and clinician predictions; ML vs clinician prescribing). The case for a conversational signal is strongest where the decision itself is socially driven: outpatient pediatric visits, typically for respiratory or diarrheal complaints, where most antibiotics are not indicated.

The closest existing work is COSTA (Wang et al.). It hand-codes the structure of consultation transcripts and finds three things: caregivers resist the doctor's recommendation in 41% of visits, the treatment phase has the longest and most contested sequences, and physician resistance to advocacy cuts prescribing odds by 77% (OR 0.23). COSTA works on text, after the fact, and describes the treatment phase at the moment the negotiation happens. We differ in three ways. We model the continuous acoustic and timing channel rather than categorical action codes. We use only the part of the visit before treatment is raised. And we test whether that signal adds information beyond both the words and COSTA's own advocacy code. Our hypothesis is that a physician who is already accommodating the caregiver's voice during history-taking is signaling that they perceive the pressure. That perception is the step that comes just before capitulation.

Research Program

The method draws on three sources from the adaptation literature, each in a different role. The main measurement tool is the Contextual Entrainment Distance of Lahiri et al.. It uses self-attention within each speaker's turns and cross-attention across speakers to score how strongly one speaker's turn is shaped by the other's. The model is trained with a self-supervised task, telling real consecutive turns from shuffled ones, on which it reports 92.13% accuracy on the Fisher telephone corpus. Xiao et al. provide a simpler, interpretable directional measure: they project the therapist's MFCC and pitch features into a space derived from the patient's speech, weight frames near turn boundaries, and compare the two with a symmetric KL divergence.

Two findings constrain the design. Both Lahiri et al. and Xiao et al. find that the professional-to-client direction is the informative one. That asymmetry becomes a testable prediction here, not an assumption. Lahiri et al. also report that models built on general adult conversation transfer poorly to clinical dyads. For that reason we do not use a Fisher-trained model as is. Because the shuffled-turn task needs no labels, we retrain it directly on the consultation audio.

The third source is McKernan et al., who show that latency between turns within a topic is an automatable behavioral marker. We add physician latency after caregiver turns as a timing feature. We also track cross-speaker attention entropy, following McNeill and Levitan, who found that this entropy falls sharply at topic boundaries. We use it to locate the approach to the treatment phase automatically.

Concretely, the pipeline runs as follows. It diarizes and aligns each visit. It uses the COSTA phase labels to cut the audio at the start of the treatment-recommendation phase. It then computes turn-level physician-to-caregiver and caregiver-to-physician entrainment trajectories, plus physician latency and prosody summaries, over the portion before treatment. A small regularized classifier maps these to the outcome.

The outcome is an antibiotic prescription in a visit where guideline review, done by two pediatricians blind to the audio, judges antibiotics not indicated. A secondary outcome is any antibiotic prescription. The data would be the audio of the 318 video-recorded COSTA consultations, which already carry phase, action, advocacy and prescribing codes. This requires collaboration with the COSTA authors and fresh ethics approval.

Three hundred visits is small; the Bone study had 28 children. We therefore pre-register the feature set, use nested cross-validation grouped by physician, and plan a prospective replication corpus collected under the same coding scheme in a second hospital system. The replication is the confirmatory test.

Baselines

The naive reference is the base rate of inappropriate prescribing, with each visit also given the physician's own historical prescribing rate. Some doctors simply prescribe more, and a sensor that only recognizes the doctor would look good for the wrong reason. Beating this means an area under the ROC curve (AUC) clearly above the physician-rate model within physicians, not only across them.

The strongest existing predictor is COSTA's manual caregiver-advocacy code, restricted to advocacy that occurs before the treatment phase. It carries the largest reported effect in the file: OR 9.23 for advocacy and prescribing, and OR 0.23 for physician resistance (Wang et al.). We found no reported AUC on Lacuna for predicting prescribing from conversation, so we will compute this baseline's AUC ourselves on the same folds. Beating it requires the acoustic model plus the code to outperform the code alone, with the bootstrap confidence interval on the AUC difference excluding zero.

The strongest automatic competitor is a text model given the automatically recognized transcript of the same pre-treatment prefix. It is a fine-tuned Chinese clinical BERT-style classifier and a prompted large language model, both given the same folds and the same hyperparameter search budget as the acoustic model. Krishna et al. show that treatment plans are recoverable from dialogue text and that an extract-then-summarize pipeline is robust to 10% simulated word error, so this is a serious baseline. The structured-visit analogue of the bloodstream-infection XGBoost models covers child age, symptoms, fever, duration and visit type. It provides the tabular reference. Its reported 0.68–0.74 pre-species AUC (ML vs clinician prescribing) is for a different task and is not a target.

Two ablations isolate the key ingredient. The first compares physician-to-caregiver entrainment against caregiver-to-physician entrainment and against each speaker's prosody alone, which tests whether the adaptation, not the voice, carries the signal. The second recomputes entrainment with caregiver turns taken from a different visit. That preserves each speaker's style but destroys the coupling, and should erase the gain if our claim is right.

Why not just…?

Why not just use the transcript with a language model? This is the real threat. Explicit requests such as "can you give him something stronger" will be caught by text, and COSTA's large advocacy effect suggests much of the signal is lexical. Our bet rests on COSTA's own observation that resistance is often indirect and is carried by sequence structure. It also rests on the autism result that clinician acoustics outpredicted lexical child features (Bone et al.). The text-prefix baseline and the stacked text-plus-audio model settle it. If audio adds nothing over text, we will report that the sensor is redundant, and the recommendation becomes a text-only prompt.

Why not just ask the caregiver about expectations before the visit? A waiting-room question is cheap. It probably captures stated demand, but not the physician's perception of pressure, which our hypothesis names as the proximal cause. The prospective corpus will include a one-item expectation question as an additional baseline. If it matches the acoustic model, it should win on cost, and we will say so.

Why not just rely on guidelines and prescribing alerts? Alerts act at order entry, after the recommendation has been voiced and negotiated. Even in hospital bloodstream infections with guidelines in place, clinicians over-treated 44% of patients and under-treated 30% (Machine learning and clinician predictions). Our contribution is timing: a signal available before the treatment phase. Whether an early prompt changes behavior is a later trial, not a claim of this note.

Why not just predict resistance from EHR data? That answers a different question, which drug will work, and it already hits a ceiling for microbiological reasons. Our target is whether any antibiotic is prescribed when none is needed. Tabular data cannot see the caregiver's pressure at all.

Decisive Checks

The idea survives only if four things hold on held-out physicians and then on the prospective replication. The pre-treatment acoustic model must beat the physician-rate reference. Adding it to the advocacy code and the text prefix must raise AUC by a confidence interval that excludes zero. The physician-to-caregiver direction must carry more signal than the reverse. And the gain must vanish under the mismatched-caregiver control.

A useful secondary analysis tracks predictive AUC as a function of how far into history-taking the audio extends. It shows how early the signal appears, which determines whether a prompt could arrive in time.

If the physician-to-caregiver direction is not the informative one, or if text explains everything, the "adaptation as sensor" idea does not carry over to stewardship. That result would also be worth reporting.