The diagnosis of canine hypothyroidism can sometimes be challenging, requiring a comprehensive evaluation of clinical signs, clinicopathological abnormalities, and thyroid function tests. Consequently, the decision whether to treat or not, or to suggest further diagnostic testing, could be complicated. Predictive models could support clinicians in managing suspected hypothyroid dogs.

The aim of this cross-sectional study was to develop a prediction tool to assist the decision making in clinical practice.

The electronic database of the Veterinary Teaching Hospital was searched for dogs tested for hypothyroidism between January 2006 and June 2020. Hypothyroidism was diagnosed in dogs with compatible clinical signs and thyroid function tests (i.e., low serum total thyroxine [T4] plus high serum thyrotropin [cTSH] concentrations or suggestive recombinant human TSH stimulation test). Dogs were excluded if medical records were incomplete or if a clear-cut distinction between hypothyroidism or euthyrodism was not possible. Eighty-two hypothyroid dogs and 233 dogs where hypothyroidism was suspected but then excluded, were included.

After data cleaning and comparison between the two groups, presence/absence of dermatological signs, serum concentrations of cholesterol, T4, and cTSH, haematocrit, were identified as relevant variables to insert in the dataset. Cholesterol, haematocrit, T4 and cTSH were expressed both as quantitative and qualitative variables and combined with dermatological signs in four different models: reduced qualitative model (RQ), extended qualitive model (EQ), reduced quali-quantitative model (RQQ) and extended quali-quantitive model (EQQ). The extended models included serum T4 and cTSH concentrations, while reduced models did not. For each model different machine learning algorithms (CART classification tree, Random Forest, Gradient Boosting Machines [GBM], Support Vector Machine, Naive Bayes, Generalized Linear Model, K-Nearest Neighbor) were applied to assess the predictive performance. Each model was evaluated and internally validated by mean of bootstrapped training-test procedure: 500 times the dataset was randomly divided into a training-set and a test-set on which the predictive capacity was calculated and reported as area under the ROC curve (AUROC). The implemented procedure yielded excellent performances, with different algorithms producing different results in the four models: the best performances were provided by the naive Bayes in the RQ model (AUROC=0.847; 95% confidence interval [95% CI]=0.843–0.851]) and by the GBM in both quali-quantitative models (AUROC=0.987; 95% CI=0.986–0.988). A beta version of the software including the four models is currently undergoing an external validation phase.

Based on our results, implementation of this prediction tool in clinical practice could prove useful, particularly in primary care practice.

Disclosures

Federico Fracassi; Financial support: Dechra, MSD, Monge; Speaking and consultancies: Boehringer Ingelheim, Dechra, MSD, Royal Canin, Hill’s, Nestlé Purina, La Vallonea.