## Beginning with a Pond of Fish That “Appear Normal”: Why Visible Symptoms Are Often a Lagging Indicator During an early-morning pond inspection, the fish suddenly exhibit reduced feeding activity. Several individuals swim close to the pond perimeter and occasionally gasp at the surface, yet the Dissolved Oxygen (DO) reading has not clearly exceeded any threshold, and no characteristic ulcers are visible on the body surface. At this point, opinions on site often diverge rapidly. Some attribute the problem to oxygen depletion, others suspect irritation caused by Total Ammonia Nitrogen (TAN), while still others recommend treatment for a possible parasitic or bacterial infection. Should the operator exchange water, suspend feeding, increase aeration, disinfect the pond, submit samples for laboratory analysis, or administer medication immediately? Every option carries a cost, but delaying action may be equally consequential. In some acute disease outbreaks, mortality may begin within **24–48 hours** after abnormal swimming behaviour first becomes apparent (Pakingking et al., 2011). This is precisely where the difficulty lies: the fish are already exhibiting signs of distress, but the available evidence remains incomplete. This is not an isolated challenge encountered by only a small number of farms. Global aquaculture is estimated to incur approximately **\$6 billion** in losses each year as a result of disease outbreaks. For operators, the value of dead fish recorded on the balance sheet represents only part of the total loss. Reduced feed intake slows growth, while deteriorating feed conversion ratios increase the cost per unit of production. Delayed grading and harvest disrupt cash flow, and emergency water exchange, diagnostic testing, medication, and additional labour further compress profit margins. Even when the stock ultimately survives, uneven growth during recovery, secondary infections, and persistent pathogen carriage may transfer risk to subsequent production cycles. High-density ponds, industrial aquaculture facilities, and Recirculating Aquaculture Systems (RAS) are particularly susceptible to the amplification of such pressures. **Both high stocking densities and poor water quality increase disease risk.** As biological loading rises, the consequences of declining DO, accumulating metabolic waste, and fluctuations in microbial communities are amplified concurrently. High stocking density does not, of course, inevitably lead to disease. The critical question is whether oxygenation, filtration, waste removal, feeding management, and biosecurity capacity have been upgraded accordingly. If even one component fails to keep pace, a localised anomaly may rapidly develop into population-wide stress. On-site management is not a standardised exercise in identifying a disease from an image; rather, it requires decisions to be made under severe time constraints using incomplete and sometimes contradictory information. Many current AI-based fish-disease products reduce this process to a simple workflow: “upload a photograph of the fish—receive a suspected diagnosis.” Such tools are undoubtedly convenient for preliminary screening, knowledge retrieval, and remote communication. However, treating a single image as the endpoint of diagnosis can easily lead users to overestimate their capabilities. Darkened body colour, anorexia, and abnormal swimming may occur during viral infections (Pakingking et al., 2011). Surface gasping may result from environmental hypoxia, gill damage, parasite burden, or multiple secondary infections. Visible symptoms often emerge relatively late in the course of disease. A photograph cannot reveal changes in water quality and provides no information on population-level behaviour, feeding records, mortality trajectories, recent operational interventions, or treatment history. What the model identifies is often a **correlated phenotype rather than the underlying cause of disease**. More importantly, early warning signals may have emerged considerably earlier but remained undetected because they were not recorded continuously on site. A study conducted in a RAS found that Atlantic salmon exposed to hydrogen sulphide stress exhibited increased swimming speed, disorganised trajectories, and a loss of coordinated schooling structure. These behavioural responses intensified with increasing concentrations and became more pronounced within a range of approximately **30–40 micrograms per litre**; the toxicity threshold adopted in the study was **60 micrograms per litre of $H_2S$** (Ciani et al., 2024). In other words, video analytics and machine vision may be able to detect anomalies in swimming speed, movement trajectories, and schooling structure before the toxicity threshold defined in the study is reached. However, an important qualification is required.
