Affirmed Developmental Concerns in Infant Notes as Early Autism Signals

Autism is usually diagnosed between ages three and five. By then the period of greatest brain plasticity, the first two years of life, has largely passed. Yet the medical record already holds clues much earlier. A large study of 29,929 infants in the Duke health system found that babies later diagnosed with autism had more physical therapy visits, ophthalmology visits, and intubations before their first birthday (Health system utilization before age 1 among children later diagnosed with autism or ADHD). That study looked only at billing and encounter codes, and it estimated population-level odds rather than predicting risk for individual children. The pediatrician's and therapist's notes from those same visits went unread. Those notes are where a clinician writes "poor visual tracking," "low tone, parent worried," or "does not respond to name."

We propose to read those first-year notes. The aim is to measure how much earlier, and at what screening operating point, free text can flag infants who are later diagnosed with autism, compared with the structured codes alone. The main obstacle is a double imbalance. Only about 1% of infants are later diagnosed, and even in those children nearly every sentence describes normal growth. Training an end-to-end text classifier directly on later-diagnosis labels asks a few hundred positive patients to locate a handful of informative sentences among tens of thousands. Our central design choice is to split the problem in two. First, we train a sentence-level classifier on densely labeled snippets to decide whether a developmental concern is actually being asserted. Second, we fit a small, low-capacity patient-level model on time-stamped counts of those affirmed concerns. In this design the rare outcome label supervises only a handful of parameters.

Core Claim

The sentence-level task is not rare once candidate snippets are filtered. Among sentences that mention eye contact, tone, feeding, or referral, a meaningful fraction contain an affirmed concern. The patient-level task, by contrast, is rare but low-dimensional. Our claim is that this factorization handles the compound imbalance better than resampling or end-to-end sequence models. We further claim that the resulting affirmed-concern features improve recall at a fixed high-specificity threshold over the structured utilization signatures, at observation cutoffs of 6, 9, and 12 months. The work would be new in three respects: it is the first individual-level predictive evaluation of first-year free text for later neurodevelopmental diagnosis, it reports a lead-time analysis with explicit leakage controls, and it tests directly whether text helps the comorbid autism-plus-ADHD group. That group is diagnosed with autism 1.4 years later on average than children with autism alone (Health system utilization before age 1 among children later diagnosed with autism or ADHD).

How This Differs From Closest Work

The sentence-level step reuses the method of Efficient Diagnosis Assignment Using Unstructured Clinical Notes. That system, HyDE, uses regular expressions to find candidate mentions with high recall. It then masks the condition word and trains a biomedical BERT model on the surrounding 16–64 words to decide whether the mention is affirmed, negated, or historical. It reports an area under the precision-recall curve (AUPRC) of 0.85 on hypertension and 0.82–0.95 zero-shot on unseen adult conditions. HyDE cannot be applied as-is to infants, because the target diagnosis name never appears in first-year notes. We replace disease-name rules with rules for developmental concern categories. These categories are motor tone and asymmetry, visual tracking and eye contact, social responsiveness, feeding and sensory aversion, sleep and irritability, and therapy referral. The last two Duke signatures, physical therapy and ophthalmology, motivate including referral and vision explicitly. We also add pediatric assertion types that HyDE never needed, chiefly parent-reported versus clinician-observed concerns, and "monitor at next visit" hedges.

The patient-level step departs from the longitudinal text models in the file. Clinical XLNet runs a bidirectional LSTM over embeddings of sequential ICU notes, and its sequence model beats mean pooling. However, it covers a 48-hour window with plentiful positives. Dementia-R1 uses reinforcement learning with intermediate rewards from numeric cognitive scores such as the MMSE and CDR. Infant primary-care notes have no such scores, so its training signal is unavailable here. Leveraging Natural Language Processing and Clinical Notes for Dementia Detection fine-tuned ClinicalBioBERT on truncated recent notes. Its accuracy fell from 0.814 in the binary setting to 0.381 in the three-class setting, and it discarded the earlier history where onset signs live. That paper is a cautionary result: end-to-end transformers on small, imbalanced, truncated records struggle with borderline cases. Our factorization is designed to avoid exactly this failure.

Two papers motivate rather than supply the method. Inference from Structred and Unstructured Electronic Medical Data for Early Dementia Detection found that bag-of-words text alone reached 75% accuracy versus 65% for non-cognitive structured data, and 82% when the two were combined. This is evidence that notes carry signal codes miss, though it comes from a curated adult cohort. Medication Usage Record-Based Predictive Modeling of Neurodevelopmental Abnormality in Infants under One Year used SMOTE (a method that creates synthetic minority examples by interpolating between real ones) on tabular prenatal data. It reached an AUC of 0.821 for predicting a 12-month developmental screening result. That endpoint is a questionnaire, not a diagnosis, and the model uses no text.

Research Program

The cohort design mirrors the Duke study. It includes infants with at least two well-child visits before age one, linked to later diagnoses of autism, ADHD, both, or neither, with a later-respiratory-infection group as a utilization-matched control. The file lists no public infant-note corpus, so this work requires institutional EHR access to a DUHS-like cohort. We state this as a dependency rather than hide it. Notes are restricted to those written strictly before each observation cutoff.

Leakage control is part of the method, not an afterthought. Any note mentioning autism, ADHD, developmental evaluation referral, or early-intervention enrollment is either masked or excluded. We use the same masking HyDE uses for target conditions. Family history of autism in a sibling is a legitimate risk factor, not leakage. We therefore report it as a separate feature so its contribution is visible and does not inflate the text signal.

The concern-assertion classifier is trained on several thousand annotated snippets drawn from rule hits. Sampling is stratified so that hits from future-positive and future-negative infants are both represented. Annotators are blind to outcome. We pretrain the assertion head on the 2010 i2b2/VA assertion data, where HyDE reports AUPRCs as high as 0.99, and then adapt it to pediatric snippets.

The patient model is a sparse logistic regression on counts of affirmed concerns:

$$\operatorname{logit} p_i = \beta_0 + \sum_{c}\sum_{w} \beta_{c,w}\, n_{i,c,w} + \gamma^\top s_i$$

Here $p_i$ is infant $i$'s risk of later autism. The count $n_{i,c,w}$ is the number of affirmed concerns of category $c$ recorded in age window $w$, with windows of 0–3, 3–6, 6–9, and 9–12 months. The vector $s_i$ holds the structured utilization features from the Duke study. The total number of parameters is on the order of a hundred, which a few hundred positives can support. Positive notes are not oversampled at any stage.

Baselines

The naive reference is a constant predictor. Its AUPRC equals prevalence, about 0.0115 for autism in the Duke cohort. Any useful model must clear this by a wide margin, and we report AUPRC lift over prevalence alongside raw AUPRC.

The structured-only baseline turns the Duke findings into a predictor. It is the same sparse logistic regression using only $s_i$: counts of procedures, intubation, physical therapy, ophthalmology, emergency and inpatient encounters, and birth stay length, adjusted for race, sex, ethnicity, and insurance. The Duke paper reports only adjusted odds ratios, such as AOR 3.5 for physical therapy and AOR 3.1 for ophthalmology. We found no reported result on Lacuna for individual-level prediction of autism from first-year structured utilization. This baseline is the comparison that answers the central question.

The strongest text competitor is TF-IDF features from all pre-cutoff notes, with leakage masking applied identically, fed to an L1-penalized logistic regression with class weighting. This follows the bag-of-words approach that reached 75% accuracy in Inference from Structred and Unstructured Electronic Medical Data for Early Dementia Detection. It is the most dangerous competitor because L1 penalties already cope with few positives and many features. A second text competitor is a Clinical XLNet-style BiLSTM over note embeddings, trained with a class-weighted loss and, separately, with random oversampling of positive patients. A third is ClinicalBioBERT fine-tuned on the most recent concept-dense notes, following the Melbourne setup. We found no reported result on Lacuna for any text model predicting later autism from infant notes. The nearest cited numbers are on other tasks: HyDE's 0.85 AUPRC for adult hypertension mention classification, and an AUC of 0.821 for 12-month screening abnormality in the PKUBC-T birth cohort. Neither is directly comparable to our task.

For fairness, every model sees the same notes, cutoffs, masking, and patient-level splits. Splits are by time of birth, so later-born infants form the test set. All models get the same hyperparameter search budget of 50 configurations, tuned on a validation fold by AUPRC.

Beating the baselines means two things. First, a statistically significant gain in sensitivity at 90% specificity over both the structured-only model and TF-IDF-plus-codes, under paired bootstrap. Second, a gain that holds at the 9-month cutoff, not only at 12 months.

Two ablations isolate the key ingredient. The first replaces the assertion classifier with raw rule-hit counts, which tests whether distinguishing affirmed from negated or normal mentions matters. The second feeds the same concern snippets end-to-end into a multiple-instance model, one that learns from bags of snippets labeled only at the patient level and pools the top-scoring snippets. This tests whether factorization, rather than snippet selection alone, is what helps.

Why not just…?

Why not just use structured utilization codes? They are cheap and portable, and the Duke associations are strong. But the codes record that a physical therapy visit happened, not what the therapist saw. Codes may also partly reflect neonatal complications that are not specific to autism. Bag-of-words text beat non-cognitive structured data (75% versus 65%) in Inference from Structred and Unstructured Electronic Medical Data for Early Dementia Detection, but that result is from adult dementia. The structured-only baseline settles the question directly here. If text adds nothing at 90% specificity, that is a publishable negative result about the limits of first-year notes.

Why not just run class-weighted L1 logistic regression on TF-IDF of all notes? This is a real threat, and it may match us on AUPRC at the 12-month cutoff. Our hypothesis is that it degrades more at earlier cutoffs. Bag-of-words cannot tell "parents deny feeding problems" from "feeding problems." With fewer notes per infant, negated and routine screening language dominates the counts. We will report the gap as a function of cutoff. If TF-IDF matches the factorized model at all cutoffs, we will conclude that the assertion classifier's value is interpretability and auditability rather than accuracy, and say so.

Why not just oversample positive patients, or generate synthetic positive notes, and train a sequence model end-to-end? Duplicating a few hundred positive patients lets a large model memorize their particular phrasing. SMOTE does not apply to text sequences, as its tabular use in the PKUBC-T study illustrates. Synthetic narratives risk changing assertion context, for example turning a hedge into an affirmation. Both the oversampled BiLSTM arm and the multiple-instance ablation test this directly. The decisive diagnostic is the gap between training and test AUPRC, alongside test performance.

Decisive Checks

We report AUPRC, sensitivity at 90% specificity, and calibration slope at cutoffs of 6, 9, and 12 months. We also report lead time, defined as the median months between an infant first crossing the risk threshold and the eventual diagnosis.

Three outcomes would confirm the proposal. First, the factorized model beats structured-only and TF-IDF-plus-codes at 9 months. Second, the raw-count ablation loses most of that gain. Third, the text gain is at least as large in the comorbid autism-plus-ADHD subgroup, the group whose later diagnosis suggests acute encounters mask developmental signs.

Two outcomes would falsify it. One is that all gains vanish once sibling family history is separated out. The other is that they vanish under a temporal split, which would indicate the model learned site- or era-specific documentation habits rather than infant behavior. We also report results for ADHD and for the respiratory-infection controls. A model that flags heavy utilizers in general, rather than developmental divergence, would score the respiratory-infection group as high-risk, and we would see it.